USER_CHANGELOG.mdView source

Documentation status (5.8 section added 2026-11-02, the release in preparation): Source-verified for OpenIntelligence v5.0 on August 25, 2026. Entries added or revised this pass are checked against the code and commit that implements them; the vector-loss and self-heal fixes are additionally confirmed by real device console traces captured the same day, not simulator or code reading alone. Anything device-only that has not been traced, including Private Cloud Compute behaviour, is build-verified rather than device-verified and is called out where it matters.

OpenIntelligence User-Facing Changelog

This document provides a chronological history of user-facing changes, highlighting how OpenIntelligence continuously improves transparency, speed, and reliability.


v5.8 - November 21, 2026

5.8 is about answers you can trust and an app that gets out of your way. A new search model and a new reranker find the right passage more often in testing, tables come through whole, every citation opens on its exact words, and the app is steadier from launch to answer.

Getting Started

  • Small tips show you around, when you need them. A short tip with a small arrow points out anything that isn’t plain from the screen, like where an answer came from, what Deep Think does, or what an icon-only button is for. It appears at the moment it’s useful, for example “Not in your documents” right under an answer that found nothing to stand on. Only one shows at a time. Apart from a few you need first, like the one in an empty chat, none shows in the app’s first minute and a new one waits at least five minutes, no more than two show each time you open the app, the tour ends after three days, and each goes away once you’ve used it, closed it or seen it three times. Settings › How This Works › Show Me Around brings them all back. Four tips carry a button that does the thing for you (Try a Question, Show Me, Go Deeper, New Library), and Settings › How This Works has a new card, Where Things Are, that lists where each part of the app lives.
  • The guides don’t use your plan. The app’s three guides now sit in their own OpenIntelligence Guide library, which counts toward neither your documents nor your libraries, so a free plan’s five documents and one library are all yours. Ask the guide library anything about how the app works.
  • Settings and the Database tab look like one app. The Settings list opens with a status panel showing your model, answer mode, plan and libraries, and each row has a small dark icon tile. The Database tab opens with the same panel, picks libraries and sections with small chips, and sets every card, heading and number in one style. Every tab now has the same small title at the top.
  • The rest of the app matches. The Documents tab, Library Settings and the sheets they open, the Atlas, and the chat’s toolbar and empty state use the same cards and text sizes, and no text in them is smaller than Apple’s minimum. Library Settings now starts with what you can change and what it means; the library’s analysis, its search model and its diagnostics show with Technical Details on (Settings, General).
  • Calmer Settings panels, and glass where things float. The panels at the top of Settings pages are plain cards with color only on each figure’s symbol and bar. The cards that float over the app (search setup, the import queue, chat notices, the setup checklist) and the chat’s text field use Apple’s Liquid Glass. In a row of figures, every block and tile now takes the row’s height, and the Database panel’s captions read “this library” instead of a library name cut short.
  • Every device gets the same careful answers. Deep Think aimed lower and search considered fewer passages on some iPhones and iPads than on a Mac, by a device table that was out of date. Every supported device now uses the settings the answer tests measured, and only the pacing follows your device. A MacBook shows its real battery, and an M6 Mac or a MacBook Neo is recognised for what it is.
  • On Maximum, the fast form is prepared after search is ready. The model that puts the best passages first is prepared once after each update. On Maximum it used to be prepared in its fastest-running form before search could start. Search now starts on the standard form and switches to the fast one in the background once it is ready. The first setup still takes a minute or two on an iPhone.
  • Search tries again if its setup fails. If the model that picks the best passages could not be set up, search used a rougher ranking until the app was opened again. It now sets the model up again before the next question.
  • Settings has fewer, fuller pages. Apple Intelligence and Private Cloud Compute are one page, Model & Privacy, where you can also choose where answers are written (On-Device, Hybrid or PCC), the same choice as the model button in the chat. It opens on a panel of what reaches the cloud and how much each model reads at once, and where answers are written and when Private Cloud Compute may be used are rows of buttons like the Intelligence Mode ones. Appearance and Notifications are one page, General. The list itself is regrouped: the settings come first (Intelligence; Libraries & Plan; This Device; App, with General and Siri & Shortcuts), then Learn (How This Works, Plain English, Version History, About), then Advanced. Technical Details is now the one switch for the technical version everywhere, Plain English included, and keeps its setting. Siri & Shortcuts lists the phrases to say, every action folded under them and, on iPhone and iPad, one Open in Shortcuts button. Advanced uses the same cards as the other pages, and About no longer repeats Version History, Device & Performance or the support card at the end of Settings.
  • Model Parameters shows only what changes an answer. The controls Apple’s model never used are gone: the penalties, the system prompt, the presets, response length and context length. The context window your device reports is shown instead. Temperature and sampling say they apply to Standard answers, which cap temperature at 0.4; Deep Think and Maximum set their own.
  • Model Parameters shows where your answers are written, and can put its settings back. The Active Model card now names the model your device runs and where the chat writes your next answer, read from the chat’s model button; it said Cloud while On-Device was picked. Restore recommended settings puts temperature, sampling and Same wording each time back to what a new install uses, and a line says which ones you changed. Your Own Reasoning Profile has four examples to tap and a Clear button, and says it’s sent only with Deep Think and Maximum answers written on Private Cloud Compute, where a test found no measurable reasoning at any level.
  • The Database tab says what each part is for, and what each repair does. Each section opens with one line on what it shows. Maintenance starts with the health of both search indexes, and Check, Rebuild and Reclaim each say what they change, how long they take and whether anything is removed. A repair is marked as the next step only when a check finds a reason, and the two rebuilds ask first. The Search section’s tips now describe what its search does: every word you type must appear, and OR, NOT and quotes aren’t supported there.

On the Mac

  • The Mac app works like a Mac app. The window has a sidebar: Chat, Documents, Atlas and Database at the top, your libraries under them, and the open library’s conversations under those. Click a library to open it, or right-click it for Library Settings, where it’s kept (iCloud or this Mac only) and Delete. Each area keeps its place while you’re in another, so an answer goes on being written while you look at your documents.
  • Settings is its own window. Choose OpenIntelligence › Settings or press Command-comma. Its pages are listed in a sidebar, and it opens on the page you last had open.
  • Documents is a table. Sort it by name, kind, passages, words, pages, size or date, select several documents, and press Delete to remove them. A document’s details open beside the table, and Add, Paste, Semantic Search and Library Settings are in the toolbar.
  • Answers and conversations. An answer’s details open beside the chat instead of over it. Right-click a message for Copy, Share, Show Details, Ask Again or Go Deeper. Share opens the Mac’s share menu, Export Conversation saves through a Save panel, and Attach opens the file panel straight away.
  • Database, the Atlas and the menus. Database picks its section with a segmented control and the libraries it counts from the toolbar, and its cards use more of the window. The Atlas’s full-screen map opens in a window of its own. Go › Atlas is Command-3 and Go › Database is Command-4, and the Import Queue moved to Option-Command-3. iPhone and iPad are unchanged.

Answers

  • Measured against 5.7. On 17 mixed test questions asked the same way in both versions, Standard got 15 right where 5.7 got 9, Deep Think 14 where 5.7 got 11, and Maximum 14 where 5.7 got 9. One run each, in the iOS 27 Simulator with Apple’s model.

  • Tokens per second means the answer’s writing speed. The speed in the bar under an answer read 1 to 5 while the answer was plainly being written faster, because it counted the time spent finding passages and reading them, and on Deep Think and Maximum every step of the reasoning, as writing time. It now counts only the answer: its tokens from the moment its first words appear to its last, and on Deep Think and Maximum only the final answer. While the answer is being found the bar shows ”–”. Answers from before this update keep the figure they were saved with.

  • The badge under an answer agrees with its Fact Check. A Deep Think answer could show a red Unverified while its Fact Check said 5 of 6 claims were verified. The check that failed was wrong: after the new search model arrived, syncing put iCloud’s copy of a library’s old search vectors back over every rebuild, so an answer looked far from its own sources and search by meaning was working from the old model’s vectors. Sync now keeps the newer copy. A library that still holds old vectors has those documents rebuilt once, and if that doesn’t fix them it shows its Rebuild button; until then that check is marked as not run and its answers read Partially Verified. When that check is the only one that fails and some claims are verified, the badge reads Partially Verified and the answer’s details say what didn’t check out.

  • The model your device runs, and the window it reports. Apple’s page on its on-device models (“Detecting available Apple Foundation Models”, read November 17, 2026) names two. AFM 3 Core has a window of 4K to 8K tokens, which can vary even within one version of the model. AFM 3 Core Advanced has 8K and runs on iPhone Air, iPhone 17 Pro, iPhone 18 Pro, iPhone Duo, iPads with M4 or later, Macs with M3 or later and Apple Vision Pro with M5. The system picks which one a device runs; an app can’t. On my devices: my M3 Pro Mac runs AFM 3 Core Advanced at 8,192 tokens, and my iPhone 16 Pro Max, on AFM 3 Core, read 4,096 on iOS 27.0.1 and 8,192 on iOS 27.2.1. Model & Privacy names the model your device runs, and the Glossary no longer gives any one device’s figure, because it changes.

  • A continued answer runs on into the window, not 1,600 more characters. When an answer runs long, the app asks Apple’s model to continue it. Each continuation used to stop after 1,600 more characters on every device, wherever the cut fell, so an answer could end mid-sentence. Each continuation now runs on into the room the window leaves, counted with Apple’s own tokenizer, up to three times.

  • Deep Think reads more of your documents where the window is 8K. Its last step, the one that writes the answer, used to read at most 3,200 tokens of passages and notes on every device. Where the model’s window is 8,192 tokens it now reads up to about 6,000. On 53 table questions asked both ways in the iOS 27 Simulator with Apple’s model at an 8,192 window, it got 49 right where the old limit got 47, and none that the old limit got right went wrong. One run each. A device with a 4,096 window reads the same as before.

  • An answer Apple pauses finishes when you come back, or says it stopped. On my iPhone a Standard answer stopped 23 seconds after I left the app, when Apple paused the model, and it stayed half-written with nothing to say so. Now, if Apple pauses the model once part of an answer has come in, the app asks the model to carry on from where it stopped: on iPhone and iPad it first waits until you’re back (up to ten minutes), and on a Mac it tries again after 3, 10 and 30 seconds. If that fails, or the model stops for another reason, the answer ends with a line in italics saying it stopped early and to ask again.

  • The best passage stays in the answer. Two extra searches look for tables and cross-references. They used to put what they found ahead of the passage the search ranked best, and the answer could then be written from the wrong document while the right table was left out. What they find now goes right behind it. And a table full of numbers was mistaken for a table of contents and pushed down; tables are now judged as tables.

  • No more blank answers. If Apple’s model stops after a stray number in a Standard answer, the app writes the answer again, and if that fails too it says it couldn’t find the answer instead of showing nothing. A one-number answer such as “60” is no longer erased. A “how much” question gets an amount, not a count.

  • “Getting search ready”, once. The first launch after an update spends about a minute preparing search on the Neural Engine. A card at the top of every screen now says so, with a bar, and that a question asked meanwhile waits for it. Leave the app and it keeps going for a while, shown in the Dynamic Island and on the Lock Screen.

  • Compute Profile, lit up. The GPU Acceleration setting is now Compute Profile: your chip’s CPU, Neural Engine and GPU light up with how much of the work each one takes, and each step from Efficiency to Full Power marks what it adds.

  • Each Compute Profile step now does a little more. Efficiency does half as much at once, so imports go slower and load the device less while reading just as carefully; Balanced is unchanged; Performance searches deeper (more passages reranked, more keyword matches, more hard spots on a page re-read); Full Power goes deepest. No step reads or searches less than Balanced. Settings › Device & Performance › Compute Profile

  • The top Compute Profile step is now called Full Power. It was called Maximum, the same name as the Intelligence Mode that reasons longest, and the two were easy to mix up. Your saved setting carries over. Settings › Device & Performance › Compute Profile

  • Your device, as Apple and the hardware describe it. The Active Device card now shows your model, chip, memory, CPU cores, GPU memory and live thermals, and under each value where it came from. Until now it showed a Neural Engine figure on every device, and Apple publishes one for only twelve chips (A12, A12X, A14, A15, A16, M1, M1 Ultra, M2, M2 Pro, M2 Max, M2 Ultra and M4), so on every other device the figure was an estimate. The tile now shows the Neural Engine’s core count from Apple’s spec pages on every Apple Intelligence chip, with Apple’s TOPS figure beside it only where Apple published one. It also said “Active cooling” on an iPad Pro, which has no fan; showed a fixed “Thermal: High” in place of the real state; rounded memory to a size the device never reported; named the iPhone 18 Pro Max’s and iPhone Air’s chips wrongly; and took four iPad models for other iPads. A Mac never showed its model name at all. A Mac now shows its model name, from Apple’s identification pages, for every Mac the app lists, M1 to M6, and the MacBook Neo. About, the readout under an answer and the setup checklist now show the same facts. Settings › Device & Performance

  • A bigger Mac reads more scanned pages at once. How many scanned pages a Mac reads at once, and the size of the GPU’s buffer cache, came from chip lists that stopped at M4 or A19, so an M4 Max got what a base M4 got and a chip newer than the list got the least. A Mac now reads one page per performance core, at most one per 2 GB of memory and 16 at once, and never fewer than its chip generation had; an M3 Pro keeps 6. The GPU’s cache takes its size from what Metal says the Mac’s GPU can use. iPhone and iPad keep their values. This follows from the Mac’s own figures and hasn’t been timed on an M4 or later Mac. A fanless MacBook Air also gets the heat limit for reading pages that was meant for it, which the app never applied on a Mac.

  • Apple’s names, and no Neural Engine claims Apple doesn’t make. The model card shows Apple’s own name for the model, read from the device, instead of “AFM 3 Ecosystem”, and image creation is called Image Playground instead of “ADM 3”. The answer route’s badge says On instead of NE, and “Running on this device” instead of “Running locally on Neural Engine”. The search model’s details say Core ML picks the Neural Engine, GPU or CPU on each device. A device without Apple Intelligence is told Apple’s own list, which includes the MacBook Neo, the iPad mini with A17 Pro or A20 Pro, iPhone Air and iPhone Duo.

  • Intelligence Mode shows what each mode really does. The page now matches Compute Profile: pick Standard, Deep Think or Maximum on a ladder, see what each step adds, and see a dot under every mode that runs each step. How Much answers the same six questions in every mode, in the same words: how much the model holds on this device, and in Private Cloud Compute when you allow it; how much it reads per step; how many steps; when it stops; and how many passages beside each match it adds. It names the on-device model Apple says your device runs (AFM 3 Core, or AFM 3 Core Advanced on a device that has it). The old list named four techniques that no mode runs (HyDE, contextual compression, extractive summarization and cross-document entity linking). It also listed steps that only Standard runs under Deep Think and Maximum, and counted import steps as mode features. Those are gone or in their right place. Settings › Intelligence Mode

  • Intelligence Mode and Compute Profile hold still. Switching modes made the page jump, because each mode’s wording took a different number of lines. Each part now keeps the height of its longest wording, so nothing below it moves. The names under each ladder are pills, so they read as buttons, and Device & Performance puts Compute Profile right under your device. Settings › Intelligence Mode, Device & Performance

  • Device & Performance and Intelligence Mode look and read the same. Every section has the same title row and badge, and every figure uses the same tile with where it came from underneath. Intelligence Mode’s panel gives search, reasoning and checking their own colours, the way the chip panel does for the CPU, Neural Engine and GPU. Active Device keeps the hardware facts and now shows the GPU family Metal reports; the live temperature moved to Live System Monitor, which uses the same tiles. Compute Profile no longer repeats a sentence under GPU in the background. Settings › Device & Performance, Intelligence Mode

  • Plan & Usage, Private Cloud Compute and Libraries & iCloud get the same panel. Plan & Usage shows how many of your documents and libraries you’ve used and how many Maximum runs are left today. Private Cloud Compute shows how much your device reads at once beside how much Apple’s servers do. Libraries & iCloud shows how many libraries stay on this device, how many sync and how many watch a folder.

  • X-Ray labels you can turn off. The X-Ray now says which part of the chip is working (ANE for the Neural Engine, GPU or CPU) inside its outline, in that part’s color, and names the Taptic Engine in its own. A Labels switch in Settings, General hides the names.

  • Live Activities and widgets in your color. Imports, long answers and search setup in the Dynamic Island and on the Lock Screen, and the Home Screen widgets, now use the app color you pick in Settings, General.

  • A see-through X-Ray, out of the way. The hardware X-Ray’s legend now lets the screen behind it show through, and it fades aside while a menu, a sheet or the “Getting search ready” card is up, then fades back.

  • The X-Ray shows exactly where your chip is. On every iPhone from the 15 Pro to the 18 Pro Max, and on recent iPad Pro, iPad Air and iPad mini models, the outline sits on the chip itself, measured from teardown photos fitted onto Apple’s repair images and X-ray scans. It’s a thin line that appears as the chip works and fades when it rests. A device whose chip position couldn’t be measured shows no chip outline rather than a guess. On iPad it stays put when you turn the screen.

  • New Chat stays new. A new chat now opens again after you close and reopen the app, instead of your last conversation.

  • Library Settings tells the truth. Each search style shows its real numbers, the model card shows the model your library really uses, and settings that haven’t been tested say so.

  • Answers find the right passage more often. The step that re-scores each search’s best candidates now uses IBM’s Granite reranker. On 246 research-paper questions, the passage holding the answer made the top five 48% of the time with the model the app ships, up from 42%, measured on a Mac. It’s a bigger model, about 68 MB more, so each search takes longer; with the old search model’s second copy gone, the app’s models still take about 20 MB less space than in 5.7. The first launch after installing spends a while setting it up on your device’s Neural Engine: the model pill says Preparing search while it does. It runs on your device like everything else, and unlike the old one it wasn’t trained on data licensed for research only.

  • Quotes only really quotes. Pick Quotes only and the answer is the sentences from your documents, word for word, each with where it came from. No model writes anything, and nothing leaves your device. Through 5.7 Apple’s model wrote those answers. Deep Think and Maximum need a model, so Quotes only answers in Standard.

  • A Deep Think answer you leave is asked again. If iOS closes the app while a Deep Think answer waits for you, the next time you open the app the question is asked again on its own. An answer kept on your device also waits for you instead of calling the model in the background, where iOS slows it down.

  • Deep Think’s panel shows what ran. It counts this answer’s model calls, not the last one’s, marks a skipped step instead of ticking it, and leaves out timings that were never measured.

  • The model is rebuilt only when a setting it uses changes. Saving a setting again with the value it already had no longer starts the model over.

  • One cleaner card while an answer runs. An answer kept on your device shows a single card on the Lock Screen and in the Dynamic Island, with its step, its progress and a small Stop. It says Paused when it is waiting for you.

  • No more crash on coming back to the app. Picking the suggested questions for a large library could freeze the app long enough for iOS to close it. That work now runs in the background.

  • A long answer pauses when you leave and finishes when you come back. iOS limits how much an app in the background can use Apple’s on-device model, so a Deep Think answer used to stall there and then be thrown away. Now it pauses at the step it reached and picks up from there when you open the app. If iOS closes the app while it waits, a Deep Think question is asked again the next time you open the app; a Maximum answer is lost and you ask again. The Deep Think icon at the top of the screen goes away as soon as an answer stops.

  • A smaller Silicon HUD on iPhone. The hardware overlay at the top of the screen is three lines instead of six on iPhone: the chip and free memory, then each part of the chip with its live reading. iPad and Mac keep the full one.

  • Settings opens on the first tap. Rows like How This Works and Plain English could take several taps to open. They open on one now.

  • Taps no longer get lost after iCloud catches up. Each time iCloud brought changes in, the app re-read your libraries in a way that froze the screen for up to several seconds, so a tap in Settings or Chat could do nothing. That reading now happens in the background.

  • Every citation shows its passage in the document. Tapping a citation used to highlight the passage only in a PDF. Now any document opens at the passage, highlighted, with a line saying whether it was found word for word, found in part, or not found. For Word, Excel and the like you see the text the app read from the file, not the original page. It shows the words are in your document; whether the answer read them correctly is yours to check.

  • Citations land on the exact words. A citation now finds its passage even where the file prints a different dash, quote or spacing than the app stored, and marks a table’s cells one by one. The passage is drawn round in the app’s color, on scanned pages and photos too. Across 44 PDFs, the share of passages marked over most of their length went from 58% to 89%, measured on a Mac.

  • Illustrate under an answer, and source chips that open. Illustrate now appears with an answer’s other actions where Image Playground is available, and tapping a source chip opens that passage in its document.

  • Deep Think checks every claim it lists. Its Fact Check list read Not Checked for every claim, even under a Verified answer. Each claim is now checked against its sources, as Standard does, and the answer itself is unchanged. If none of the claims holds up, the badge says so instead of Verified.

  • A number stays with what it belongs to. On my iPhone, Standard said its similarity floor was 0.20 and Fact Check called that supported, because 0.20 was in the passage it cited. It was Maximum’s, on the next line of the same list; Standard’s is 0.28. The app had handed the model the list’s numbers without the names they belong to; now each line it passes on keeps its name. And a number counts in Fact Check only where the passage gives it to what the answer says it belongs to, in the same list item, line or section, so a claim that borrows another one’s number is marked unsupported. Tap a citation and its page now says the passage was found, and whether the answer’s number is given there for what the answer says.

  • A “no” about something else isn’t a contradiction. Standard answered 0.28 right, and Fact Check still said 0 of 1 verified, because the guide’s line also says “no query expansion” and the check read that “no” as the passage disagreeing. Now a “no” only counts against an answer when it’s about something the answer says.

  • Deep Think searches side by side. The five searches behind a Deep Think or Maximum answer used to run one after another. Now the first runs, then the rest two at a time, and the answer draws on exactly the same passages.

  • iPhone stays upright. Turned sideways, the iPhone app squeezed the chat into half the screen. It now stays in portrait on iPhone. iPad and Mac are unchanged.

  • Answers have to answer the question. Every check used to ask only whether an answer was in your documents, so a page number copied from a paper could pass as the answer to “how much”. Now Apple’s on-device model also checks that the answer gives what you asked. In Standard, an answer that doesn’t is written once more. A value pulled from a table no longer replaces an answer it doesn’t fit. An answer that still doesn’t answer is never marked Verified.

  • Answer details tell the same story as the badge. The details card called almost every answer Source-Locked, “all claims fully verified”, even when its checks had failed or no document was used. It now shows the same result as the badge under the answer, and says Not Checked when no check ran.

  • Tables reach the answer whole. A question about a table could be answered from a short summary that showed only its first two rows, so it could give the wrong row’s value. Now the whole table goes to the model, in every mode and in libraries you already have. Files you import from now on also stop taking a journal’s page header as a table’s title.

  • A PDF is stored as it reads. Where a PDF carries its own text, the app read many pages by recognizing the page image instead, so a misread like “1-28” for “1-2 g” or “TU” for “IU” became what search and answers saw. Now the text the PDF itself holds is used wherever it matches what was recognized, in tables and in the rest of the page, and running headers, page numbers and journal lines stay out of your passages. Word, PowerPoint and Excel files also keep hyphens, symbols, superscripts and table cells they used to drop or shift. Measured on eight research papers, the numbers kept went from 90% to 94%, and misread numbers fell by half. Files you imported before keep their old text until you import them again.

  • Library Settings describes what it does. It said documents were split by comparing sentence meanings and that the library would re-chunk itself; neither happens. It now says files are split at headings, tables and sentence ends, and that rebuilding is always your choice.

  • Chat no longer turns white. Chat could show a back arrow where the new-chat button belongs, and tapping the Chat tab while you were already in Chat could leave a white screen. The cause was the tips: each one attached to a button in Chat’s top bar left an invisible screen of Apple’s on top of the chat. Tips now attach only while they’re showing, and if Chat ever disappears while you’re on its tab, it builds itself again within a second.

  • Maximum answers from your documents’ own words. Maximum’s last step used to write from the short notes its reading sessions kept, so a number could come out as a note had rounded it: it once gave “$1.2 billion” for a figure the document states as $1,577 million. It now reads the source passages themselves at that step. On 30 questions asking for a value from a research paper’s table, it gave the right value on 25, up from 8 without the change, and faster at the median (79 seconds against 117). Measured in the iOS 27 Simulator.

  • Deep Think opens with the answer. Its answer now starts with one sentence that answers the question, with its source, before the detail. On 53 table questions it was right on 46 with this and 45 without, and about 14 seconds faster at the median, in the iOS 27 Simulator.

  • Needs Verification means a claim failed. A Deep Think answer could carry the Needs Verification banner even when every claim it listed was supported. Now it appears only when a claim failed its check. On my 20 test questions the banner left five right answers and stayed on the wrong one.

  • “Which document” gets the document’s name. Asked which document mentions something, an answer used to say “[S1]” and leave you to tap it. It now names the file next to the citation.

  • Deep Think writes from the passages themselves. Its final step used to read only the short notes its reading passes wrote, cut at 3,000 characters. It now reads the passages those notes cited as well.

  • Apple’s model connects as soon as it’s ready. If Apple Intelligence wasn’t ready when the app opened, on a new device or while it downloads after an iOS update, every answer until you reopened the app was excerpts only, with no model writing. Now the app checks before each question and when you come back to it, and connects the model as soon as it’s ready.

  • Every answer carries its citations. A Standard answer that streamed in without citation numbers kept none; on my iPhone that was 8 of 28. It now gets them from its own passages.

  • The same question finds the same evidence. Deep Think and Maximum could gather different passages when you asked the same question twice, and three other steps could differ from one launch to the next. They now come out the same each time.

  • A summary comes from the documents you asked about. “Summarize my lease” could put another document’s summary first. A summary question now takes summaries only from the documents its search found.

  • The match score reads the same everywhere. The ring under an answer and the panel’s “best match” showed one result two ways, 89% and 81%. They now show the same number.

  • No totals the documents don’t state. Asked how many participants received a treatment across several studies, an answer could open with a total it added up itself, and add it wrong. Now it gives each study’s number as the document writes it, with its citation, and a headline number that no passage states is taken off.

  • A cleaner answer bar. The session badge no longer sits on top of “Thinking”, a finished Maximum answer shows one confidence figure, not two, and the line that says what the app is doing reads as plain words, without dashes or token counts.

  • An answer streams inside its bubble. The text used to come in bare and move into the grey bubble when it finished. It now writes inside that bubble from the first word.

  • Labels in a list read as headings. When an answer groups its bullets under bold labels, like “What stays on the device:”, each label now shows as a heading over its bullets, in the same style as the answer’s other headings, instead of as one more bullet at the same size. The bullets under it, bold words inside a sentence, numbered lists, tables and citations look as they did, and Copy and Share give the answer as it was written.

  • Swipe in from the left to open Evidence Threads. On iPhone and iPad, a pull in from the chat’s left edge opens your threads, as the sidebar button does. A scroll that starts at the edge still scrolls.

  • On-Device, Hybrid and PCC are one click on the Mac. They sat in a submenu of the chat’s model menu. They’re its own items now, with a check on the one in use.

  • The chat’s counts match its pickers. The document and passage counts beside the library and mode pickers have the same circled icon, glass and height, and show the exact number, like 10,000, up to a million.

  • A follow-up answers the question you just asked. In a longer conversation the answer check could judge, and then rewrite, the answer against an earlier question. It reads the one you asked.

  • An answer that found nothing no longer steers the next one. When an answer says your documents don’t cover something, in any mode, the next question in that chat doesn’t build on it, including when Apple’s model said so in its own words. A follow-up to an answer Siri gave you keeps that answer, with conversation memory on or off.

  • Change a library’s documents and the chat starts from them. If you re-imported a corrected document, the open chat could keep refusing until you cleared it, because earlier turns, refusals included, went into the next answer. Turns from before a library’s documents last changed are no longer given to the model. New Chat now also clears the library’s conversation memory, and Clear Chat stays cleared with iCloud, which used to copy the old memory back.

  • A paper’s own reference numbers stay out of your answer. When a passage held citations like [3] or [124], an answer could copy them, and [3] then looked like a link to source 3. They’re removed, and a bare [2] the model writes for a source becomes a working [S2].

  • “Getting search ready” says which step it’s on. It shows Step 1 of 2 or Step 2 of 2 with the seconds so far, where it used to sit near the end and then finish.

  • The Atlas map reads cleanly. Lines that ran from a cluster to nothing are gone: each cluster’s name now sits beside its own dots with a short line to one of them, and a name with no room is hidden along with its line. File names joined with hyphens or underscores read as words (“RAG Technical Architecture”), long names are shortened in whole words, and the words on the map are a label’s size, never drawn over each other or over the box in the corner. The Insights card no longer says a document “fits well with your library” when everything in it came in at once.

Libraries

  • Imports keep going when you leave the app. An import you left running could stop at its last step, because the search model asked for the graphics chip that iOS doesn’t give an app in the background, and coming back it sat at Queued until you left and returned again. It now switches to the processor and Neural Engine when you leave and back when you return. Starting an import while another was finishing in the background could also freeze the app until iOS closed it; that’s fixed too.
  • An answer you leave still finishes. An answer that was still being checked when you left the app could sit at “Checking sources…” for two minutes and stay there after you came back, for the same reason: the search model wanted the graphics chip. Answers now switch to the processor and Neural Engine when you leave, as imports do.
  • Long answers stay smooth. A chat with long answers could freeze the app as you came back to it; each answer’s formatting is now worked out once.
  • Leaving the app no longer ends it. When iOS takes back an answer’s background time, the app now hands it back properly, so it is paused rather than closed and the answer finishes when you return instead of starting over.
  • Deep Think keeps the answer it wrote. A last research step could replace a finished Deep Think answer with a short, wrong one; it no longer can.
  • An answer the app finished is not lost. If iOS closes the app in the background after an answer is written but before it is checked, the reopened chat shows that answer, marked as not yet checked, instead of asking your question again.
  • An answer keeps working when you leave. A Standard answer you leave the app during keeps searching and finishes in the background when Apple Intelligence allows it, retrying for about a minute before it waits for you to come back. Deep Think and Maximum still wait for you, since they make many calls to the model.
  • A right answer in your question’s words reads Verified. “Chicken should reach 165°F” was marked Partial because Fact Check counted “should” and “reach”, which came from your question and not the guide. Words you asked with no longer count against an answer; a number or a name still has to be in the passage.
  • An answer you leave reads all its evidence. An answer that paused while you were away and finished when you came back could be written from a sliver of its passage, because the app read the model’s room as zero while it was in the background. Asked “What temperature should chicken reach?”, it said the manual didn’t say; it now gets the whole passage and answers 165°F.
  • Siri never waits forever. A question asked through Siri while the app was in the background could be held up by Apple’s model and then wait for the app to come back, which it never did, while the progress bar kept creeping. Siri now offers to continue in the app, and otherwise says it couldn’t answer.
  • Your first library is Library 1. It was called General. Now that the guides have a library of their own, your first library is Library 1 and the next is Library 2, on every device. A name you gave it yourself stays.
  • The Guide library stays as it came. The OpenIntelligence Guide library holds the app’s own guides, so its documents have no Delete in their menu or on a swipe, Remove All Documents isn’t offered there, and Siri and Shortcuts don’t delete them either. On the Documents screen Add, Paste and dropping files are off there too; a file sent to it from Share, Siri or Shortcuts goes to your own first library. The Guide library itself can’t be deleted either, its settings show what it is and change nothing, and if it went missing on a device before this version it comes back the next time you open Documents.
  • A new search model, and a one-time rebuild. Search now finds passages with IBM’s Granite embedding model, licensed for commercial use, in place of MiniLM, whose training data was licensed for research only. In my tests in the app it also put a passage from the right document in the top five more often: on 15 of 19 Standard test questions, up from 10 (iOS 27 Simulator). Each library rebuilds its search index once, by itself, when you open it after the update. It keeps the text it already read, so nothing is imported again, but a large library can take a while. On a Mac the rebuild keeps the Mac awake, so it’s the best place to let it run; on iPhone and iPad it carries on in the background. If you share libraries over iCloud, update every device. If you share a library over iCloud, the device that rebuilds it first does the work: your other devices check the index they receive and skip their own rebuild when it’s already the new one.
  • Table column names stay with their values. A table whose title runs across its top row had the title taken for its column names, so its values lost the names of their columns. The title now joins the table’s caption and the real column names stay. Import a file again to fix a table already in a library.
  • Tables stay in one piece, and no words are left behind. A table whose value wrapped onto a second line, followed by a label line, could be split in two, and words a page scan didn’t recognise were dropped even when the PDF carried them as text, which lost whole rows of numbers in one paper and dozens on its reference pages. On the eight research papers I test import with, the app now passes 439 of 515 checks, up from 344 when this work began.
  • Reading page images no longer holds up the app. That wait could make iOS close the app during a long import in the background.
  • Watch a folder. On Pro and Lifetime, Library Settings has Watch a Folder. New files in it import by themselves, and a changed file replaces the version it imported, so an old answer may no longer open that version. Files you delete from the folder stay in the library, and your originals are never changed. On a Mac it checks every few minutes while the app is open. On iPhone and iPad it checks when you open the app. It has its own row on the Documents page, under the buttons, so you don’t have to open Library Settings to find it.
  • Several files go in without waiting on each other. Each file used to wait for its overview and suggested questions before the next one started. Now the files go in first and those come right after, the same as before.
  • Importing a very large file twice no longer makes a second copy. Large PDFs and text files now get the same duplicate check as everything else.
  • A library’s count matches what’s in it. A library synced with iCloud could show more documents on its button than it held.
  • Files stay in the library you chose. A file added to a library that was deleted in the meantime no longer lands in the library you have open; it says the library is gone. Leftover records of deleted libraries are cleared.
  • Two windows share one engine. On a Mac or an iPad, each window ran its own copy of the app’s engine and iCloud sync, so two windows could rebuild or sync the same library at the same moment. They now share one, so switching libraries in one window switches it in the other.
  • The Atlas shows each thing once, and only what’s real. Its topics were listed four times over, some counts were of the 6,000 passages it samples rather than your library, and its similarity heatmap was drawn from random colors, so it said nothing about your documents. The heatmap and a “coming soon” placeholder are gone, the topics are one list with how much of the library each holds, the counts are the library’s own, and a topic named after a long title no longer ends mid-phrase.
  • The app no longer gets closed in the background during iCloud sync. Sync did its file work on the app’s main thread, so while iCloud held one of its files, the app stopped responding, and iOS closed it when that lasted 10 seconds. Sync now does that work in the background.
  • On a Mac, the app no longer rewrites its iCloud sync files when nothing has changed. Every iCloud sync pass wrote two of its sync files again, changed or not, and each write looked to the app like a new change from iCloud, so it synced and reloaded your libraries again about every 5 seconds. On my Mac that was 470 to 715 reloads an hour, with the app at 46% CPU and Spotlight and iCloud at about 30% each while I wasn’t using it. A pass that changes nothing now writes nothing.
  • You can feel the app working. An answer pulses as it searches, as Deep Think works through each step, and as the text writes, and it taps once for each check that passes. Each file taps as it goes in. Each tap plays one haptic instead of two, and the app no longer buzzes on its own when the battery, the charger or the phone’s temperature changes.
  • The app no longer stops responding at launch while iCloud catches up. With libraries synced through iCloud, the window could freeze for a few seconds at launch while iCloud finished with one of the sync files. The app now waits for iCloud in the background.
  • No freeze at launch from a watched folder. With a watched folder, the app could freeze for up to eight seconds at launch while it looked the folder up. That now happens in the background.
  • Finished imports no longer stop a library from syncing. Imports that had finished, or whose files were gone, could stay in the shared import list, and while they did, iCloud left that library’s search index out of every sync. They’re cleared at launch now.
  • Rebuilding a library leaves your other devices alone. A rebuild used to take each document out and put it back as a new copy, so another device signed in to iCloud took the old one as deleted, cleared that library and rebuilt it too, back and forth, and a document’s file could be lost along the way. A rebuild now keeps every document as it is, and the library shows how far it has got, or that it’s waiting for another library to finish first. Like an import, it shows on the Lock Screen and asks to keep going when you leave the app, and if the app is closed it starts again next time.
  • A finished or paused import’s card leaves the Lock Screen and your Watch when you open the app. It used to be able to stay there for good.
  • Every file you import gets its suggested questions. If you switched libraries or tabs, or left the app, while the questions were being written, the run stopped and was saved as “no questions”, so Chat said none were ready and a file could go without any for good. A stopped run now saves nothing, and Chat writes questions for any file that has none.
  • An import keeps showing progress while it reads a PDF. The progress bar used to sit still for the whole time a PDF’s pages were being read, so with the app in the background iOS could take it for stuck and end it. It now moves page by page.
  • The import log is in color. Each step of an import has its own color in the processing log (reading, splitting, embedding, indexing, saving, analyzing), so you can tell them apart at a glance.
  • An import reads Complete when you can search it. The import card waited for the overview, tags and suggested questions too; on my iPhone a 27-page PDF was searchable 46 seconds before the card said so. It now says Complete as soon as the document can be searched, and the rest follows.
  • Footnotes stay out of the middle of a sentence. A page’s footnotes go after its body text, and a sentence that a caption, figure or table interrupted is put back together.
  • Sync never deletes a document it can’t place. A file no library record names is moved aside for 30 days instead of deleted, and sync leaves every file alone while an import or rebuild is still running.
  • The Atlas opens at once. Its analysis of a library took 24 to 61 seconds on 6,000 passages and ran again every time you opened it. It’s now kept on your device and shown straight away, and redone only when the library changes. An axis name that doesn’t fit at the end of its axis is dimmed, with a small arrow pointing to it.
  • Shared links open the app again. My website moved to gunndamental.com, and links copied with Copy Link pointed at the old address, which opened the web page instead of the app. Links now use the new address.
  • Atlas names come from your documents, not the app’s labels. Clusters and axes could be named “Path & Section” or “Table & Headers”, words the app itself adds when it stores a passage. Those are left out, labels that would sit on top of each other give way, the legend lists every cluster, and counts read “16,554 passages”.
  • Faded scans and fine print. A scanned page is now also read in four overlapping quarters, which picks up small print the full-page pass missed, and a washed-out scan has its contrast raised before it’s read.
  • Text in another script stays. A Greek, Hebrew, Arabic, Cyrillic, Chinese, Japanese, Korean or Hindi line inside an English document was thrown away as scanning noise. It’s kept.
  • Spreadsheets, saved web pages and photos. A CSV saved by Windows software, or one with a quote mark inside a value like a 55” screen, imports correctly. A saved .html file imports as readable text with its tables, not as raw markup. A portrait photo is read the right way up; it used to be read on its side.
  • Libraries & iCloud says where your libraries are kept. Settings › Libraries & iCloud replaces “Storage location” with a plain description of where your libraries live, with the technical details below it.
  • The Mac’s menu bar item asks any library. It shows the app’s own icon now, and with more than one library it has a picker, so you can ask a library by name without opening the app first.
  • Deleting a library deletes what the app remembered of its conversations. The questions and answers kept for a library’s conversation memory stayed on your device after the library was deleted. They go with it now.

Plans

  • The plans screen shows the plan you have. If your Lifetime came from a purchase before 5.5, the plans screen said Free and offered you plans you already had, while Plan & Usage said Lifetime. It says Lifetime now.
  • The plans screen says what each plan adds. “Why upgrade now?” is “What you get” now, with what paying actually changes.
  • Free, Pro and Lifetime, side by side. Under the three plans there’s now a table of everything a paid plan changes and what each plan has of it: Maximum mode, documents, libraries, conversations in each library and watched folders. Your own plan is marked. The conversation limit (5 on Free, 20 on Pro, unlimited on Lifetime) was always there and wasn’t written anywhere.
  • Plans are one tap from Settings. The Plan & Usage row opens the plans screen itself, with what you have used at the top, and the Plan block at the top of Settings opens it too. The screen is shorter: your plan and usage, the three plans with their prices, the side-by-side table, then restore, manage, redeem and the legal links. The banners, the “What you get” list, the slides and the tip are gone; the table and the plans said the same things.
  • Annual first, and fixes across Settings. The plans screen lists Pro (Annual) first, shows the “Data Not Collected” privacy label right under your plan, and your iPhone taps once when a purchase or restore goes through. Settings no longer shows the Developer page, About states the right minimum system (iOS 26 or macOS 26), the top of Settings agrees with the Private Cloud Compute page, small text in the panels is at Apple’s minimum size, Intelligence Mode’s step names wrap instead of being cut off, and a library with no documents in the Database tab says so, with a button to add some.

Ratings

  • Rating the app is one sheet, Apple’s own. The “Enjoying OpenIntelligence?” alert after a thumbs-up is gone. After a good answer, or your third answer, the app asks with Apple’s rating sheet: tap the stars, add a review if you want, and submit without leaving the app. It asks once per update at most, and Apple limits it to three times a year. Write a Review in Settings is still there.

v5.7 - October 22, 2026

This one’s about getting things in and out of the app. JSON and JSONL files import as records, Siri and Shortcuts can ask, search and add to your libraries, Spotlight and Visual Intelligence find your documents, and an answer you share takes its sources with it.

Your Documents

  • A new library filled by one very large file doesn’t ask to be rebuilt. When the first thing in a library was a PDF over 10 MB or a big text or JSONL file, the library asked to rebuild its search index the first time you asked it something, right after it had been built. It doesn’t now.
  • JSON and JSONL files import as records. A .json, .jsonl or .ndjson file used to come in as one long run of raw text, and a .jsonl file couldn’t even be chosen in the file picker. Now each record is its own passage, with its field names and its numbers kept exactly as the file has them, so a chat export, a log or a dataset can be searched and cited one record at a time. A line that isn’t valid JSON is kept the way it was written.
  • OpenIntelligence shows up under “Open in”. A PDF, a text or Markdown file, a CSV, a JSON file, a Word, Excel or PowerPoint file or an image can be sent to the app from Files, Mail or a share sheet, and on the Mac by dropping it on the app’s icon. It shows you the file first: tap it to look through it, choose the library it goes to, then Add, or Don’t Add and nothing is kept. A file you add goes through the same import check as a file you pick in the app.
  • Long passages are stored the way your document wrote them. When a passage was too long to index in one piece, the app split it and re-typed it on the way, so “$75.50” could be stored as “$75. 50” and a question mark could become a period. Now it’s cut at line and sentence ends and nothing is re-typed. A document you already imported keeps the old text until it’s imported again.
  • A long passage isn’t cut short before it’s indexed. In a large PDF or text file, a passage a little too long for the search model was cut to its first 1,500 characters before it was indexed, which is about two thirds of what the model can read. The text itself was always stored whole. Now it’s only cut where the model’s limit really is.
  • Paste into a library. The Documents screen has a Paste button. A file you copied is copied in, a web address is downloaded the same way “Save a Web Page” does it, and text is saved as a note. A copied PDF or picture works too. Whatever you paste goes through the same import check as a file you pick. On iPhone and iPad, iOS may ask you to allow the paste when what you copied came from another app.
  • Share to OpenIntelligence. On iPhone and iPad, OpenIntelligence is in the share sheet for a web page or text you’ve selected. Tap Add, and the next time you open the app it’s added to the library you have open: the page is downloaded and saved the way “Save a Web Page” does it, and text is saved as a note. A file still goes in with Open in.
  • Big text files don’t freeze the app anymore. A big text file used to freeze the app while it imported, sometimes for minutes, and iOS could close the app while it was stuck. The step that cuts a file into passages was re-reading the rest of the file for every passage, and it ran on the same thread as the screen. Now it reads each part once and runs in the background, the import shows how far along it is, and Cancel works while it runs.
  • Very large text and JSONL files import a piece at a time. A text or JSONL file used to be read, cleaned up and cut into passages all at once. An 88 MB JSONL file with 189,579 records took 2.8 GB of memory that way on an iPhone, and iOS closed the app before it was a fifth done. Now a file over 8 MB is read a few megabytes at a time, and each piece is indexed and saved before the next one starts, so the whole file is never held in memory at once. A JSONL file keeps its record numbers across the whole file, and its short records share a passage instead of each taking one.
  • An interrupted import picks up where it was. Each saved piece of a large text or JSONL file is a checkpoint. If the phone locks or you leave the app, the import continues from the last saved piece when you come back, and it doesn’t ask. If iOS closed the app, Resume does the same. The screen also stays on while an import is running, so the phone doesn’t lock itself in the middle of one.
  • iCloud sync waits for an import to finish. While a large file was importing, iCloud sync went back and read the whole library again after every piece, at least 135 times in one import of mine, and that slowed the import down. In a library that already had other documents, it could also have saved the library without the file that was still coming in. Now sync leaves a library alone until its import is done.
  • A file that’s too long says so. A text or JSONL file of any length now comes in a piece at a time (above). For other kinds of files, one document can hold 50,000 passages, and a longer one used to be cut off there without telling you. Now the import stops with a message that says to split the file into smaller ones.
  • The Atlas opens on a very large library. The Atlas used to read every passage in a library before it drew anything, and on a library with tens of thousands of passages iOS could close the app while it was working. Now a library with more than 6,000 passages is mapped from 6,000 of them, taken evenly from across the whole library, and the Atlas tells you when it’s showing a sample. Sorting those passages and laying out the 3D map, which took seconds on a large library, run in the background now instead of holding up the screen, and the Content Breakdown reads whole words, so “capital” no longer counts as technical writing because it has “api” in it.
  • The Atlas map finishes on a very large file. With one very large file in a library, the map stayed on “Mapping semantic space…” and never drew. Two of the things it shows were being worked out from every character of the file, 88 million of them in mine. Now they’re worked out from an even sample of the text, and the map draws.
  • The Atlas names things from your own documents. Cluster names used to come from a built-in list of subjects. One library of mine got three clusters that were all called “API Reference”, the names changed from one launch to the next, and the axes were labelled “Return”, “Document” and “Text”. Now a cluster is named by the words that set it apart from the rest of the library, no two clusters share a name, each axis is labelled at both ends, and the same library gives the same map every time. There’s no list of subjects behind it anymore and nothing in it is tied to English, so it works from whatever your files say. On another library of mine every name still came out as two words starting with A (“Agent & Alert”, “Annual & App”), because a word in three passages counted the same as one in all of them and ties went alphabetically. Now the words most of a cluster’s passages share come first. On my library, one end of an axis still had no words at all, and another read “Blade & Gently”. Now an end is named by words that set its passages apart even when only two of them share one, an end with none is marked “Component” instead of being left blank, an adverb like “gently” doesn’t name an end, and no word names two ends. For a library in French, German, Italian, Portuguese, Spanish, Turkish or Chinese, the app asks your device to load that language’s word classes when it doesn’t have them yet, so the same goes there.
  • The Atlas’s 3D map keeps all three axes on screen. When you turned the map, an axis could slide off the edge of the screen and take its labels with it, and every change on the page put the map back where it started. Now the map is the size it was before 5.7 and fits the width of a phone held upright, and when you turn an end of an axis off the screen, its name waits at the edge of the map nearest it, clear of the buttons and lists over the map, so every axis keeps its names in view. The map stays at the angle you leave it until you change what it shows or a setting. Each end of each axis and each cluster is named in the map, and now for what it is: the document it runs toward, or a part of one document, like a manual’s cleaning or troubleshooting pages, and by its words only where neither stands out. If you’d rather have a plain cube with the axis names in a key under it, pick Chart under Map in the Atlas’s settings. The full screen view has an All Libraries choice that shows your libraries in one map, each in its own color and named in the map, so you can see how they sit next to each other. A library made with a different embedding model can’t share that map, and the view says which ones it leaves out. The Atlas also missed anything imported while it was open until you left the tab and came back. It updates when the import finishes now, and the 3D Semantic Map tile, which didn’t open anything, opens the full screen map.
  • A very large library doesn’t hold the app up. With 144,247 passages in one library on my iPhone, the app read the whole library twice every time it opened, iCloud sync compared both copies of it passage by passage, opening Chat scored every passage to come up with suggested questions, and the first question ran a grammar step over every passage before it started. Now the library is read once, sync skips the comparison when nothing has changed, and those steps work from an even sample of a large library. The app was also measuring the whole library again every time it opened (its languages, its vocabulary, how long its sentences are), for the analysis it shows in Library Settings. Now it measures a sample of a large library, keeps the result, and only measures again when the library changes. The grammar step before the first question is kept the same way. Opening a library that size still takes a few seconds.
  • Numbers in tables are kept. A table cell that held only a small number, a year or a part code, like “34”, “2024” or “SKU10293”, was dropped when the file was imported, so a question about it had nothing to find. That happened in CSV files, spreadsheets and tables in PDFs. Those cells are kept now. A table you imported before needs importing again to get them back.
  • A deleted document or library stays deleted. With two windows open on the Mac, a document or a library you deleted in one could show up again later, and a library could also come back on its own right after you deleted it. Both stay deleted now.
  • Two switches in Developer & Diagnostics stay as you set them. Pipeline Trace and Force Reasoning Chain went back to what they were the next time the app opened, because changing one of them wasn’t saved. Both are saved as soon as you change them now.
  • You can undo a delete. A document you delete leaves the library right away, and an Undo button shows at the bottom of the screen for a few seconds. Tap it and the document’s back where it was. Otherwise it’s deleted when the seconds are up, or as soon as you leave the app.
  • Swipe a document to delete it. On iOS 27 and macOS 27 a document in the library can be swiped away. Like any delete it can be undone for a few seconds, and a document in an iCloud library still asks first.
  • Pause an import. The import queue has a Pause button now. A large file that’s saved in pieces goes on from where it stopped when you continue, and any other file starts over, and the queue says which. Waiting files wait too, until you continue or discard them.
  • Leftover text from an unfinished import is cleaned up. When the app closed in the middle of an import, the document’s text stayed in the search index with no document to go with it. Now it’s removed the next time the app opens. And when you stop an import, the app’s copy of the file goes too, so choosing the same file again keeps its name instead of getting a “-2”.
  • Text printed sideways along a page’s edge stays out. A journal’s “Downloaded from” line runs down the edge of every page of some papers, and it was read into the page’s text, where it could start a passage or become a heading. It’s left out now.
  • Addresses, phone numbers and amounts are found everywhere in a document. Since 5.3 the app has used Apple’s own detector to recognise addresses, phone numbers, dates and amounts of money, but it only ever looked in tables. Now it reads the whole document, so an address in a paragraph of a lease or a phone number in a manual’s text counts too.
  • A code in a picture is read, not guessed. When the app describes a picture in a document, or a photo you import, it now reads any QR code or barcode in it with Apple’s Vision and writes down what it holds. Before, Apple’s model was left to read the code from the picture, and in a test on my Mac it made up a link the code didn’t hold.
  • Keyword search covers every page of a large PDF. A PDF over 10 MB is read fifteen pages at a time, and each batch numbered its passages from zero. The keyword index files a passage under its document and that number, so a passage from a later batch that reused a number wasn’t stored. Search by meaning still had it, but an exact word or a part number on a later page could be missed. The numbering carries on through the whole file now. Search by page had the same kind of gap: each batch cleared the pages before it, so only the last fifteen pages of a large PDF could be found that way. Every page is kept now. A large PDF you imported before needs importing again to get those passages back into keyword search.
  • Pages in Greek, Chinese, Arabic and other scripts keep their text. The step that clears junk lines out of a PDF page only counted unaccented Latin letters and Cyrillic letters as letters, so a line in another script looked like symbols and was removed. It counts letters in every script now. And a PDF in Greek, Hebrew or another script that Apple’s text recognition doesn’t read is read from its own text, where a check that compares the page with a scan of it used to throw that text away.
  • Text files that aren’t UTF-8 come in readable. A text file saved as Shift-JIS, GB 18030, Windows-1251 or UTF-16 used to come in as rows of accented letters and symbols, because anything that wasn’t UTF-8 was read as Latin-1. Now the app uses Apple’s own detection to work out how the file was saved and reads it that way.
  • A recording’s name doesn’t pick the wrong language. The app took the language of a recording from two letters in its file name, so “notes_friday.m4a” was transcribed as French and “plans_december.m4a” as German. Now the two letters have to be the last word of the name, like “lecture_fr.m4a”, or the language has to be spelled out.
  • Recordings use Apple’s newer transcriber. The app had code for Apple’s long-form speech transcriber that was never actually built into it, so every audio or video file went through the older recognizer. Now a recording goes to the newer one first, and to the older one if the newer one can’t take it. On iOS 27 and macOS 27 the newer one also reads audio the app couldn’t open before, and when a video has more than one sound track, it tries each one until it finds speech. Both run on your device.
  • Import progress counts each document once. The progress your iPhone shows for an import that keeps going in the background could read 100% the moment one document finished, with more still to go, and then drop back. Now each document counts once, the Live Activity’s count goes up as documents finish instead of starting over, and Check Import Status says how far along the import is.
  • Pictures are described from the picture. When you import a document with figures, or a photo, the app asks Apple’s model for a sentence or two about each picture so search can find it. It turns out the picture itself was never sent, only the labels another tool had guessed for it. On iOS 27 and macOS 27 the model is shown the image now, on your device.
  • A photo with a table in it keeps the table’s words. When you imported a photo or a screenshot that the app read as a document, a table in it was saved as a block of programming text instead of its rows, so the prices on a receipt were buried in it and search and answers read that text. Now the table is saved as its rows and cells. A photo imported before 5.7 keeps the old text until you import it again.
  • New libraries on iPhone and iPad use the faster copy of the search model. The app carries two copies of the model that turns passages into something it can search by meaning. I timed both on an iPhone 16 Pro Max, the model alone. The one new libraries had been using did about 88 passages a second, and the other did about 100 to 120 depending on the setting. On the processor alone the first one managed 5 a second and the second one 84. New libraries on iPhone and iPad use the second one now. A library you already have keeps the one it was built with, because changing it would mean indexing the whole library again. Library Settings and the sample documents used to say the model ran on the Neural Engine either way. They say which copy runs where now, and the Default badge sits on the copy your device uses for a new library.
  • An import stops if the search model comes back empty. If the model that turns passages into something the app can search by meaning handed back empty results, the app stored them anyway, and the library looked fine but could only be searched by its words. Now the import stops and says so, and a library that was already built that way is rebuilt when you open it.
  • A document that went missing from search comes back. A document could stay listed in a library with nothing of it left in search, so no question could find it. One of mine had 455 passages and none of them were searchable. Opening the library now finds a document like that and rebuilds it.
  • Import a whole folder. Drop a folder on the Documents screen, or choose one when you add documents on the Mac, and the files inside it that the app can read come in, from the folders inside it too. A folder used to be refused.
  • Word counts are right. A document’s word count split words at spaces only, so a line break joined two words into one, and it added up the document’s passages, which counted the words two passages share twice. Both are fixed.
  • A file goes to the library you added it to. A file waiting in the import queue went into whichever library was open when its turn came, so switching libraries while files were waiting moved them. Each file now remembers its library.
  • An import in the background finishes. After your file was searchable, an import kept calling Apple’s model for an overview, tags and suggested questions, and saved the document partway through. In the background iOS could end it during that work: the Lock Screen said “Task failed”, and ended early enough, the document wasn’t saved. The document is saved first now, and in the background the extra work waits until you open the app.
  • No “Error Processing Document” without Apple Intelligence. On a device where Apple Intelligence isn’t available, the Documents screen opened on “Error Processing Document: Apple Intelligence is required but unavailable.” Your documents import and search without it, and the chat already says what still works, so that alert is gone.
  • An interrupted import keeps its place across an app update. When the app was updated in the middle of a long import, the import started again from page 1, because it no longer recognised its own saved pages. It recognises them now, so an update costs only the page it was on.
  • PDFs over 500 MB import. A PDF bigger than 500 MB was turned away, even though big PDFs are read a few pages at a time. Now any PDF imports, as long as there’s room for it on your device.
  • A resumed import picks up faster. When a big PDF’s import was interrupted, picking it up again re-checked every page and re-embedded every passage it had already done. Now it skips the pages it already read and the passages it already embedded.
  • An import iOS stops in the background says it’s paused. When iOS took back the time a big import had in the background, the banner said the task failed, though the import picks up again when you open the app. It says it’s paused now.

Shortcuts and Siri

  • Asking an empty library from Siri or Shortcuts says it’s empty. An Ask action pointed at a library with no documents used to get an answer from Apple’s model on its own, with nothing from your documents in it, and it came back looking like an answer about that library. Now it says the library has no documents yet.
  • Actions hand back what they found. Asking a question from Shortcuts used to show a card and give the shortcut nothing to work with. Now the ask actions return the answer with its sources, List Documents returns the documents, and List Conversations returns the conversations, so the next step of a shortcut can use them. When an action can’t do its job it stops the shortcut and says why, where it used to report success with an apology inside. “Ask About an Image” and “Search with a Photo” hand back their text as well. An answer or the passages a shortcut gets are kept on your device so the steps after it can use them, and removed the first time the app runs after they’re a day old.
  • You can pick the mode. “Ask My Documents” and “Ask a Library” have a Mode menu: Standard, Deep Think or Maximum. On the free plan, Maximum from Shortcuts counts against the same daily allowance as Maximum in the app.
  • “Ask About a Document” only reads that document. It used to put the file’s name in the question and search the whole library, so the answer could come from a different document. Now it’s searched inside the one you picked and nowhere else, and so are Summarize a Document and Compare Two Documents. These three answer in Standard.
  • Plainer names. “Query Documents” is “Ask My Documents”, “Search Library” is “Ask a Library”, “Ingest Document” is “Add a File” and “Ingest Webpage” is “Save a Web Page”.
  • “Save a Web Page” downloads the page. It used to say it was extracting the page, and nothing was ever downloaded. Now it fetches the page, saves its text and imports that. A link straight to a PDF saves the PDF. A page that only shows its text after running scripts can’t be read this way, and the action says so.
  • One action adds a file, a link or text. “Add to Library” takes any of the three, and you can pick which library it goes into.
  • Find Passages. A new action that returns the passages matching a search without writing an answer, so the next step of a shortcut can read them.
  • Create a library and switch libraries. Two new actions, “Create Library” and “Set Active Library”.
  • Set the answer mode from a shortcut. A new action, “Set Answer Mode”, switches the chat to Standard, Deep Think or Maximum, so a shortcut or a Focus automation can pick the mode before you ask. On the free plan, Maximum follows the same daily allowance as in the app.
  • Read the settings from a shortcut. A new action, “Get Settings”, hands a shortcut the answer mode and the active library, so it can change them for a task and put them back afterwards.
  • Open a library or a document. Two new actions, “Open Library” and “Open Document”. Open Document takes you to the document’s own page.
  • Siri’s answer card opens its sources. When Siri or Shortcuts shows an answer, the card has a button for each document it comes from, up to three, and tapping one opens that document. An answer from a whole library also has Ask a Follow-up, which opens the app in a new conversation that starts with that question and answer, so your next question picks up from it, and a Standard answer has Think Deeper, which asks the question again in Deep Think. Both open the library you asked about, and the chat stays in Deep Think until you switch it back.
  • Check a claim. A new action, “Check a Claim”, takes a sentence and tells you whether the words and figures in your documents hold it up, partly hold it up or don’t, and which passage it checked against. It’s the same check an answer gets, without writing an answer, and when one of them says no where the other says yes, it tells you to read the passage.
  • Delete documents from a shortcut. “Delete Documents” asks before it deletes anything, and each document’s passages and Spotlight entries go with it.
  • “Add a Document” and “Scan a Document” do what they say. They used to open the app and stop there. Now one opens the file picker and the other, on iPhone and iPad, opens Apple’s document scanner, the same one Notes uses, and adds what you scan to your library as a PDF.
  • Documents and libraries carry their details. A shortcut can read a document’s name, type, date added, page count and library, and a library’s document count and dates. Before, it got a title and nothing else.
  • A button into Shortcuts. Settings, under Shortcuts, has a button that opens the Shortcuts app on OpenIntelligence’s actions, and the list there names every action as Shortcuts shows it.
  • “Add a File” keeps its file. It queued the temporary copy that Shortcuts hands over, which the system can remove when the action ends. Now the file is copied into the app first.
  • Siri phrases are things you’d say. “Ask OpenIntelligence a question”, “Save this page to OpenIntelligence”, “Find passages in OpenIntelligence”.
  • Long questions from Shortcuts keep going. On iOS 27 and macOS 27, a question asked from a shortcut or Siri can keep running in the background past the usual 30 seconds, and the system shows its progress while it works. On iOS 26 and macOS 26 the progress shows in Shortcuts too. If it’s stopped, by the system or by you, it stops writing and says which it was.
  • Siri asks before using Private Cloud Compute. When a question you ask with Siri or Shortcuts needs Private Cloud Compute and you haven’t said yes or no to it yet, the app comes up and shows the same consent screen the chat does. Before, a question like that stayed on your device, and with the app already open it could wait on a screen that never showed. If the app can’t come up, the question stays on your device.
  • Search from Siri. On iOS 27 and macOS 27, “Search OpenIntelligence for” and what you’re looking for opens the library’s search with it and runs it.
  • Visual Intelligence finds your documents. Point the camera at a device’s label or a printed page, or take a screenshot, and Visual Intelligence lists the documents in your libraries that mention what’s in it. Tap one to open it. The text in the picture is read on your device.
  • Siri knows what’s on your screen. When a document’s page, a library or a conversation is open, the app tells Siri and Apple Intelligence which one it is. Before, nothing on screen was named to them.
  • Siri suggestions. The app tells your device which library you open and ask in, and when, so Siri and Spotlight can suggest it at that time of day, as “Open” and the library’s name. It takes some days of use before anything shows up.
  • Focus filters. In Settings, Focus, a Focus can have an OpenIntelligence filter that picks which libraries the app shows while it’s on, and the first of them opens. A Work Focus can keep your work library in front.
  • The Siri phrase shows up where you’d use it. On iPhone and iPad, the empty Chat screen shows Apple’s tip with the phrase for asking a question. Close it and it stays closed.
  • Add to Library from Shortcuts and Siri uses your plan. With the app closed, it used the free plan’s limit of five documents, so a Pro or Lifetime library of five or more refused it. It uses your plan and any document packs now.
  • Find Passages stays at 20. The Limit never goes above 20, even when a shortcut passes in a bigger number.

Spotlight

  • A Spotlight result opens what you tapped. Tapping a document or a library in Spotlight used to open the Documents tab of whatever library was active. Now a library opens that library and a document opens its own page.
  • Spotlight keeps your documents. A document’s Spotlight entry used to disappear after 30 days. Now it stays until you delete the document, a passage you find in Spotlight opens its document, and when you delete a document its passages leave Spotlight too. A document you delete on another device leaves this one’s Spotlight the next time the app opens here. A passage’s entry is titled with its document and its page, where it said “Page 3 [Chunk 13]” before, and a text file’s passages no longer say “Page 1”.
  • Turning Spotlight off clears it. Turning off Spotlight Indexing in Settings used to only stop new entries, and what the app had already put in Spotlight stayed until it expired. Now turning it off takes out everything the app put there, along with the libraries it suggested to Siri, and turning it back on lists your libraries and documents again.

Getting around

  • Handoff. What you have open in Chat or Documents on one device can be picked up on your other devices. For a library that syncs through iCloud it opens in that library, and a document’s page goes across too. The app offered Handoff before, but the other device had nothing it could open.
  • Your conversations go with your library. In a library set to iCloud Drive, conversations were saved in a folder iCloud sync never looked in, so they stayed on the device you had them on. Now they’re saved where sync looks, and the ones you already have move there the first time you open the app. Deleting one deletes it on your other devices too.
  • iCloud only sends what changed. Every sync sent each library’s chat history, transcript and memory up to iCloud again whether they’d changed or not, and when two devices did that at the same time, iCloud kept both versions as “name 2”. My iCloud folder had 500 of those. Now a file goes up only when it changed, and a library’s search files go up together.
  • Services on the Mac. Select text in any app, then choose Ask OpenIntelligence from the Services menu to put it in the chat, ready to send, or Add Text to OpenIntelligence to save it to your library as a note. Select files in Finder and choose Add to OpenIntelligence to import them.
  • The menu bar on the Mac. OpenIntelligence has a menu bar icon now. Click it to ask a question, add documents or search, and the app comes forward with it. Settings, Appearance turns it off.
  • Long answers on the Lock Screen. On iPhone and iPad, a question that’s still being answered after about ten seconds, like most Deep Think and Maximum questions, shows a Live Activity with the step it’s on, and it ends when the answer does. Its Stop button stops the answer and keeps what’s written so far.
  • Put your libraries in your own order. On iOS 27 and macOS 27, drag a library in the row of libraries on the Documents screen to move it. The order stays on that device.
  • Touch and hold the app icon. Ask, Add Document and Scan Document are on the Home Screen, under the icon.
  • A widget. Add OpenIntelligence to the Home Screen or the Lock Screen. The small one asks a question, and the medium one has Ask, Add and Scan.
  • Your Library widget. It shows the library you have open and how many documents are in it, and the medium size also shows the last question you asked there with the start of its answer. If you’ve chosen to hide private widget data, the answer stays hidden while the phone is locked.
  • Scan into Library from the Lock Screen. Add the Scan into Library control to the Lock Screen, Control Center or the Action button, and it opens OpenIntelligence’s camera without unlocking. Take the page, tap Use Photo and unlock, and it’s added to the library you have open.
  • Copy Link. Touch and hold a document or a library and tap Copy Link. The link opens it in OpenIntelligence on your iPhone, iPad or Mac, and the App Store page anywhere the app isn’t installed.
  • Keyboard shortcuts. On the Mac, and on an iPad with a keyboard: Shift-Command-N starts a new conversation, Command-O adds documents, Command-1 and Command-2 switch between Chat and Documents, Command-3 opens the Import Queue, Command-F opens the library’s search, Command-] and Command-[ go to the next and previous library, and Command-comma opens Settings.
  • Sheets on the Mac open as one page. Seven screens, the citation sheet among them, opened on the Mac as a two-column split with an empty side. They open as one page now.
  • The hardware X-Ray steps aside for sheets. It floated above everything, a sheet’s buttons included, and a tap there went to it. It hides while a sheet or an alert is up.
  • The tab bar gets out of the way on iPhone. It shrinks while you scroll a conversation or a library, the way it does in Apple’s own apps on iOS 26, and comes back when you scroll up.
  • The sample guides’ questions. A library of the sample guides is meant to show a set of questions written to walk you through the app, and an older list usually showed in its place, which the refresh button didn’t change. If you’d deleted one of the guides, the library could lose the written questions altogether, and if you’d added your own documents to it, it kept asking about the guides. Now a library with any of the guides gets the questions written for the ones it has, each with the guide and section it comes from, and refresh shuffles them. Once you add a document of your own, the questions come from what’s in the library.
  • A Siri tip on the Documents screen. On iPhone and iPad, under your list of documents there’s a tip with the phrase for Find Passages, which finds the passages in the open library that match what you say. Close it and it stays closed.
  • The whole app has one look. Colours, spacing and corners were different from screen to screen, because each screen had been styled on its own. Every screen now uses the same spacing and the same shapes, and colour only means something: green, orange and red for status, and one colour per answer mode. Text sizes, the chat bubbles and the colours of the buttons under an answer stay as they were.

Accessibility

  • VoiceOver reads a library as a button. Each library in the row of libraries is one button now, read with its name, how many documents it holds and whether it’s the one you’re in. Before, VoiceOver read its icon as “Move” and its name as plain text, and never said it could be tapped. The Document Quota reads as one line too, like “Document Quota, 2 of 5 documents”.
  • VoiceOver names every switch and slider. Several switches in Settings, the hardware X-Ray’s among them, and the sliders for a library’s passage size and overlap and for the clusters in Visualizations had no name, so VoiceOver read each one as a switch or a slider with nothing to say which. Each one is read with the name shown beside it now, and a slider says the number shown next to it.
  • Easier to tap. The microphone and the paperclip in the chat field are 44 points square, Apple’s default size for a control on iPhone and iPad and the size of the send button beside them, and “I’ll add my own documents” on the first screen is 44 points tall.
  • The hardware X-Ray reads clearly. Its numbers were faint text on a see-through box. Now they’re white and in color on a solid dark backing, they grow with your text size, and the “Drag me” hint wraps instead of running off the screen. On a new install it starts off, so it doesn’t cover the chat the first time you open it. Turn it on in Settings, Appearance, with the Telemetry HUD switch.
  • Gray text is easier to read. The small gray text on Chat, Documents and Settings was lighter than Apple’s accessibility guidelines ask for. It now uses the darker gray iOS itself switches to when Increase Contrast is on, and the text on the Documents screen grows with your text size. VoiceOver also stops finding a hidden “Upload complete” panel when nothing is uploading.
  • Colored labels are easier to read. The mode, model and count pills at the top of Chat keep their color in the icon and the background, and their words are in black or white now, because colored words on a tinted pill were hard to read. In light mode the app’s blue is a little deeper too, so white text on blue buttons reads more clearly. When they don’t fit side by side, the pills at the top of Chat and in Settings stack instead of getting cut off, and at the largest text sizes the top of the Documents screen scrolls with your documents instead of running up under the title bar. The question box’s text grows with your text size now too.
  • Answers and imports work at the largest text sizes. On iPhone and iPad at the biggest text sizes, the import panel broke words in half, cut its Continue and Discard buttons short, and left files you couldn’t scroll to, and the Fact Check line under an answer split “SUPPORTED” into pieces. Now they wrap or stack instead, and everything in the import panel scrolls. Its buttons are bigger to tap there, the names and details in your document list grow with your text size and read more clearly, and the first screen’s “See It in Action” button wraps instead of getting cut off. The colored scores in an answer’s sources, the badges on a document’s page and the privacy line on the plans screen use black or white words now, like the pills at the top of Chat, and VoiceOver reads the import totals as what they count instead of the names of their symbols.

Sharing

  • A shared answer takes its sources with it. Share used to send the answer’s text and nothing else, so the [S1] and [S2] in it pointed at nothing. Now the documents and pages it cites are listed under it, with the app’s name and its App Store link at the end.
  • Share an answer as a picture. Under an answer, Image makes a card with the answer, its sources and a QR code for the app’s App Store page, ready to post or send.
  • The share sheet shows what you’re sharing. It used to show the first line of an answer, or a bare link when you shared the app. Now it shows the app’s name and icon.
  • Export a conversation. The chat menu has Export Conversation, as Markdown to read, as JSON Lines for other tools, or as a PDF to print or send. Each answer keeps its source list. A new Export Conversation action hands a saved conversation to a shortcut as a file, in any of the three.
  • Copy works on the Mac. The Copy button under an answer showed “Copied” on the Mac and copied nothing. It copies now, and so do the three other copy buttons that had the same problem.
  • Copy takes an answer’s sources with it. The Copy button under an answer copied its text alone, so the [S1] in it pointed at nothing once you pasted it. Now the documents and pages its labels point at come with it, the same list Share sends, without the app’s name and link. An answer a shortcut passes to its next step carries the same list, and in Shortcuts it shows as its opening words instead of its formatting marks.

Answers

  • Every label in an answer opens the passage it came from. The numbers on an answer’s sources could be counted differently from the passages the answer kept, so a label like [S5] opened nothing, or opened the passage next to the right one. Now they’re counted the same way, so each label opens the passage the answer was written from, and the same goes for the source list when you copy or share it.
  • Summaries and comparisons keep the second-best passage. For a question that reads whole passages, the app moves the strongest ones to the start and the end, where the model pays the most attention. It used to move them before checking what fit, so when only some passages fit, the second-best one was the one left out. Now it picks the passages that fit first and then moves them.
  • A lookup answer keeps its own words when they match the page. For a question about one value, like a code or a capacity, the app could swap a finished answer for a piece of the page around that value, and the piece could start in the middle of a word and leave out half of what you asked. Now an answer that already gives the value from your document stays as it was written, and when a piece of the page does stand in, it’s the whole sentence.
  • Answers read the rows of a table. When the passage that answers a question was part of a table, a list or a page of labelled values, the model was given only the heading above it, so it could tell you your document didn’t mention something that was right there. Now it gets the rows themselves, and so do the checks under the answer. Your libraries don’t need to be imported again.
  • Deep Think and Maximum answers are checked. Settings said every mode checks its answer against your documents, and Deep Think and Maximum skipped the checks, so their answers said Not Checked. Now they run the same checks a Standard answer does, and the badge says Verified or Unverified. The answer stays as it was written either way.
  • Deep Think keeps its full answer. When Deep Think’s answer hedged on one point, a second search could swap the whole answer for one short line, and that line could be wrong. On a question about file types, a full answer was swapped for a line about “notes, warnings and exceptions”. Now a short line takes the answer’s place only when it agrees with what the full answer said.
  • A number pulled straight from your document has to be about your question. For a question about a single value, the app can take the value straight from a passage instead of writing an answer. It could take the wrong line’s value, like the engine oil’s 4.5 L for a question about the fuel tank. Now the line the value is on has to name what you asked about, or the model answers instead.
  • Ask out loud. On iPhone, iPad and Mac, the chat field has a microphone button. Tap it and talk, and your words show up in the field as you speak. Tap it again to stop, then send when it looks right. Dictation runs on your device.
  • A simpler chat. New installs start with a plainer chat: while an answer is on its way it says what it’s doing in words (“Searching your documents…”, “Writing the answer…”), and the stats bar, the step-by-step console, the technical badges and the token count when an answer finishes stay out of the way. Fact Check says where each point was found instead of a confidence score. Turn the technical view on with Technical Details in the chat’s ⋯ menu or in Settings, Appearance. If you’ve used the app before, your chat stays as it was until you switch.
  • See the page a source came from. Tap a citation in an answer, or the page under a Fact Check claim, and a PDF opens at that page with the passage highlighted, so you can check the answer against the page itself. In an answer’s details, a source from a PDF on this device has a Page button that does the same. Other kinds of documents open their details page, where a Preview button shows the whole file in Quick Look.
  • Finished imports and answers tidy up after themselves. The notification that an import or an answer had finished, and its card on the Lock Screen, stayed for up to ten minutes or until you swiped them away, even after you’d opened the app. Now they go as soon as you’re back in the app.
  • Fact Check quotes the right sentence. The quote under a claim in Fact Check was the first few lines of the passage, wherever the claim came from. Now it’s the sentence in that passage that matches the claim best.
  • Without Apple Intelligence, an answer says why it’s excerpts. An answer made of excerpts used to end by saying the model might be rate-limited. On a device that can’t run Apple Intelligence, or with it turned off, that isn’t the reason, and now it says the real one.
  • Settings describes Deep Think as it is. The list of what Deep Think and Maximum do named four tool functions for the model. Neither mode hands the model a tool, so the line is gone.
  • Page numbers after a blank page are right. When a PDF had a blank page, every page after it was filed under the number of the page before it, so an answer that quoted one of those pages could name the wrong page. Each page keeps its own number now. A PDF you imported before this keeps the old numbers until you import it again.
  • Sentences from your documents aren’t cut in half. Before the model writes, the app picks the sentences that match your question, and it was cutting them at every line break and after things like “p.m.” So an answer could be built from “on the 5th” without the first half of the sentence. They stay whole now.
  • Numbers are checked as numbers. The check that compares the numbers in an answer with your documents used to pass a number if its digits showed up anywhere on the page. An answer said a model had “2.5 billion” parameters, and it passed because another model in a table was named “Qwen-2.5-3B”. Now a number is compared as a number: “2.5 billion” isn’t backed up by a “2.5” that’s part of a name, and “2.5” doesn’t pass because the page says “12.5”. A sentence with a figure your documents don’t state shows as Partial in Fact Check, and the answer isn’t marked Verified. Fact Check also checks an amount written out in words, like “four hundred dollars”, against the figures in your documents.
  • Amounts, dates and units are checked too. The check now reads amounts of money, dates and measurements with Apple’s own detector, the one that already finds them in your documents when you import. An answer that gives an amount in another currency than your documents do, another day of the same month, or a number with other units, like 35 kPa where the manual says 35 psi, isn’t marked Verified.
  • Right answers aren’t marked Unverified by a lab check anymore. One of the checks behind the Verified badge is meant for questions about lab research, to keep results from cell, animal and human studies apart. It was reading pieces of everyday words as lab terms, like the “um” in “number”, the “ip” in “trip” and the “cell” in “cancellation”, so right answers about travel and hotel documents were marked Unverified, and a question about travel insurance was refused. Now it reads whole words, and it only steps in when your question is about lab research. When a check does fail, the answer’s details now say which one.
  • An answer’s title isn’t checked as a claim. Fact Check scored an answer’s own heading, like “Hotel Address in Lisbon”, as if it were something the answer claimed, so a right answer could read “1 of 2 Verified”. An answer’s title is left out of Fact Check and its count now. A point also no longer shows a gap where its citation was taken out.
  • Fact Check reads the way the answer does. When a claim was cut out of the middle of bold text, Fact Check showed the bold’s asterisks with it, a time in bold read “9:00 a.m..”, and a claim that pointed at a range of sources read “are present in through.” Now those read the way the answer has them, and a long claim quoted under Information Gaps stops at a whole word.
  • The Verified badge agrees with Fact Check. An answer could show a green Verified badge while its own Fact Check said “0 of 2 Verified”, because the badge only looked at the checks on the whole answer. Now when Fact Check can’t back any of an answer’s claims, the badge says Partially Verified or Unverified instead.
  • A blank response gets one more try. Sometimes Apple’s model returns nothing the app can read. When that happens while an answer is being written, Standard now waits two seconds and asks once more before it gives up. Deep Think and Maximum already retried one of the ways that failure shows up, and now they retry all of them.
  • A question asked again looks at your library as it is now. The app kept the search behind every question you asked, and asking the same thing again reused it, even after you’d added or deleted documents. So a new document was left out, and a deleted one could still end up in the answer. Now any change to your libraries, or an update to the app, means a fresh search.
  • Old searches are deleted. The app keeps the search behind each question you ask, with the passages it found, and it never deleted them, so a library you’d deleted still had its text in there. Now they’re deleted whenever you add or delete a document, and after an update.
  • Answers draw on more of your documents. Before the model writes, the app is supposed to pass over passages that say the same thing as one it already picked, so the model sees more than one part of your documents. That step never did anything, because the passages reached it without the data it compares, so near-copies got through. It works now.
  • Error codes count. When the app picks the sentences that match your question, it skipped every two-letter word, so in “What does E1 mean?” the E1 never counted. A short code with a letter and a number in it counts now.
  • Fact Check catches a passage that says the opposite. A claim that shared enough words with a passage saying the opposite, like “Pets are allowed” against a lease that says they aren’t, read as supported. Now it reads partly supported. And a passage that says “without” something only says the opposite when the claim names that thing: a right claim about keeping a sick child home was marked as contradicted because the school’s handbook says “without fever-reducing medicine”.
  • The best passage the search finds always reaches the answer. “What happened on September 30” found the September 30 section of a monthly log and then answered from September 24 and 25, because the step that fits passages into Apple’s model kept only sentences repeating the question’s words, and that day’s lines didn’t repeat them. The passage the search ranks first now goes in whole, and the rest of the room is filled as before.
  • A date written “30th” finds its day. Asking “What happened on September 30th” still missed that day in my notes, which call it “September 30”, because “30th” isn’t a word anywhere in them and the search ranked a different day first. A day next to a month is now searched as its number, so “September 30th”, “Sept. 30th” and “the 30th of September” all look for September 30.
  • Asking for a summary gets one. “Summarize” and “summary” were read as a request to add numbers up.
  • Excerpts say they weren’t checked. When the app couldn’t write an answer and showed excerpts instead, the badge read Partially Verified, though nothing had checked them. It reads Not Checked now.
  • An answer no longer cites a document that isn’t there. Passages left behind by a deleted document, or by an import that didn’t finish, could turn up in an answer credited to “Unknown”. Searches leave them out now, and in a library kept on this device they’re removed when the app opens.
  • An answer’s details show Apple’s token count. On iOS and macOS 27, the details under an answer list what Apple’s model spent on the whole answer: tokens in, how many of those it had cached, tokens out, and reasoning tokens when Private Cloud Compute reasoned.
  • Fact Check catches three more slips. A sentence of an answer now reads Partly Supported when it leaves out a “not” your document has (“Pets are allowed without written permission” against a lease that says they aren’t), when its day doesn’t match (“the 16th day” where the lease says the 6th), or when its words are all in your documents but no one sentence of them says it. Each one says why under the sentence.
  • Private Cloud Compute always asks first. A question typed to a library with no documents yet could go to Private Cloud Compute without asking you: with Hybrid when it was too long for your device, and every time with Private Cloud Compute picked in the model picker, even after you’d said Don’t Allow. Now it stays on your device.
  • Fact Check no longer waves through sentences it couldn’t read. When the app couldn’t pick out the key words of a sentence in an answer, it counted that sentence as backed by your documents without comparing anything. That happened most in languages like Greek. Now it checks the sentence’s ordinary words against your documents instead.
  • A saved answer opens every source it cites. A saved conversation kept only an answer’s first 12 sources, so after you reopened the app a tap on [S13] or later opened nothing, and Copy left those out. Now it keeps every source the answer cites.
  • Fact Check no longer fails an answer over a figure caption. A Deep Think comparison of two papers was marked unverified because the check read a figure’s caption as the definition of “MCI”. The same question now passes.
  • A question that didn’t get its answer can be asked again. If the app is closed or stopped in the middle of an answer, the question used to sit there with nothing under it. Now there’s an Ask again button.
  • Searches don’t pick up stray letters. Asking “compare the two documents” could add letters like PA, RE and DO to the search, because they appeared inside your words. Only whole words count now.

Settings

  • The Database tab speaks plainly. It led with “SQLite FTS5” and showed storage internals to everyone, and a card said searches were “100X” faster than a method the app no longer has, which nobody had measured. Now it says what the keyword index does in plain words, the speed figures are gone, and the storage details are there with Technical Details on. Rebuild Index says what it rebuilds too: the whole-document index, not the one that answers your questions.
  • Device & Performance speaks plainly too. The chip internals, the hardware limits and the vector-math details are there with Technical Details on, and the rest says what each mode does in plain words. Language Support used to show a fixed list of 19 languages. Now it shows the languages Apple’s model reports on your device.
  • Library Settings has a Search section. Each library can pick how it searches: Automatic, which goes by what’s in the library, or Balanced, High Precision, Technical Manual or Exploratory. Under Advanced Controls you can set the score a passage needs, how many passages at that score an answer needs, how much it leans on your exact words or on their meaning, and how varied its passages are. These controls were in the app before, hidden, and the one for how many passages an answer needs only worked with one exact setting. They all work now, and once you pick something the app leaves it alone after an import instead of choosing for you.
  • A setting you change stays changed. The app saved the setting you’d just changed with the value it had before, so the last change you made could come undone the next time the app opened. It saves the new value now.
  • Send me a crash report. On iOS 27 and macOS 27, if the app closes unexpectedly, it offers afterwards to email me a report of where it stopped. It has the app and system versions, your device model and the part of the app’s code that was running, and none of your documents, questions or answers. It only goes anywhere if you send it, and Settings, About can email or share the last one later.
  • On iOS 26, Settings says what Private Cloud Compute needs. On a device that isn’t on iOS 27 or macOS 27 yet, Settings said this copy of the app was built without Private Cloud Compute. It says Private Cloud Compute needs 27 now.
  • The privacy policy is in the app. Settings, About has a Privacy Policy row. It opens the same policy as the website, and it reads without a connection.
  • Settings says when an answer goes to Private Cloud Compute. It said only an answer too large for your device went there. With Hybrid, a Deep Think or Maximum answer drawn from more than one document does too, with your permission, and Settings, How It Works and the privacy policy say so now. The choice that allows it without asking is called Allow without asking. It was called Allow for Deep Think and Maximum, though it never stopped at those modes.
  • Settings shows the model’s real window. It said 4K everywhere. Settings and an answer’s details show the number your device’s model reports now, and the Context Window card no longer says Standard switches to three sessions on a hard question, which it doesn’t. The Glossary, the sample documents and the suggested questions said 4,096 too, and now use what your device reports, 8,192 on a Mac with macOS 27.
  • Intelligence Mode in plain words. Each mode says what it does in plain words: Maximum, for one, can make up to 50 passes over your documents and stops when its findings stop changing. The techniques behind each mode are listed with Technical Details on, in Appearance, and that list no longer shows two that Standard doesn’t use. Siri & Shortcuts shows how many shortcuts the app offers instead of “10 of 10 Registered”.
  • The GPU setting says what each choice does. Each choice in Settings, Device & Performance lists what it changes, taken from what the app runs: which chips searching and importing can use, where scanned pages are cleaned up, and where large searches run. A line under it says, for your device, whether an import that keeps going in the background can still use the GPU. It used to say things the app didn’t do, like answers staying on the Neural Engine, and it made Maximum sound like more than Performance, when the two work the same way while the app is open. Its entry in Plain English said a higher setting imports faster and uses more battery, which nothing had measured, and on my Mac a higher setting made large searches slower. It says what the setting decides now.
  • The GPU button on the import panel works. It changed its own label and nothing else. It switches the same setting as Settings now, through all four choices.
  • Deep Think’s numbers are the real ones. Settings showed 5 sessions on an iPhone 15 Pro and 13 on some Macs. Its reasoning chain stops at 8 sessions on every device, and Settings and its badge now show how sure it aims to be on yours, from 85% to 95%.
  • Your library’s model is named right on a Mac. Library Settings called a library built on a Mac with macOS 27 “Apple Natural Language”, and other screens said every library was 512-dimensional and ran on the Neural Engine. Each one now shows your library’s own model, its size and where it runs.
  • The Glossary and the bar under each answer say when Private Cloud Compute is used. They said only an answer too large for your device went there. They now match Settings: a Deep Think or Maximum answer from more than one document can go too, every answer does when you pick Private Cloud Compute in the model picker, and after Always Allow the app doesn’t ask again.

Notifications

  • A notification when it’s done. Settings has a new Notifications page with one switch. With it on, you get a notification when an import finishes, or when an answer that took more than 20 seconds is ready, if OpenIntelligence isn’t in front. It’s off until you turn it on, and it never shows your question or a file name.

Plans

  • Redeem a code. The plans screen has a Redeem a Code button. It opens Apple’s sheet for offer codes.
  • New Chat at your conversation limit keeps the conversation you had. When a plan’s limit on saved conversations was reached, New Chat emptied the screen but kept saving to the old conversation, so your next question replaced everything in it. Now the conversation stays on screen with a note about the limit, and your next question is added to it.

When your device is busy

  • The app eases off when iOS asks. iOS 27 can tell apps to use less when the device is under pressure. When it does, an import leaves a short gap between batches and Deep Think and Maximum leave a second between steps. The answer isn’t cut short. It takes a little longer.
  • Long imports stay steadier on a Mac that runs warm. On the Mac, text recognition now eases off when the Mac heats up, instead of running flat out until a fanless MacBook Air throttles.
  • The app answers your taps right after you open it. For the first several seconds after opening, the app could freeze while it checked your iCloud libraries. On my iPhone that was three freezes, about 6.7 seconds in total. The check now runs in the background, and the first screen also draws faster.
  • Starting an import doesn’t freeze the app. When an import began, the app could stop responding for two to four seconds while it got the search model and the page reader ready. That work now happens in the background.

v5.6 - October 8, 2026

This one’s about your first few answers. The citations in an answer open now, numbers and times come through the way your document wrote them, and the Verified badge only shows when a check actually ran and passed.

Answers

  • Citations open. An answer cites its sources as [S1], [S2] and so on, and tapping one did nothing, because the app only made a link out of a bare [3]. Now each one opens the passage it came from.
  • Numbers and times stay the way your document wrote them. After the model wrote an answer, a cleanup step that’s there to remove repeated sentences rebuilt every answer, repeats or not, and it split the text at every period, so “$300.00” came out with a space in the middle and “10:00 p.m.” lost its “m.” Now it leaves an answer alone unless it finds a real repeat.
  • Two facts don’t get merged into one. That same cleanup could treat “E1 is an open circuit” and “E2 is a short circuit” as one sentence said twice and drop one of them. Sentences with different numbers or codes are never treated as repeats now.
  • Answers stop ending with my own prompt. When the app wasn’t confident in what it found, some answers ended with a line like “What sources show, What’s missing, Confidence note”, which was part of my own prompt. That line’s gone, and the app asks for one plain sentence about what your documents don’t cover instead.
  • No more asterisks. The Fact Check rows showed bold text with its asterisks, and an answer showed ”###” and ”**” while it was being written. Both are cleaned up.

The Verified badge

  • Verified means a check ran and passed. The badge showed Verified for any answer that wasn’t on a short list of failures, so an answer whose checks failed, or never ran, still said Verified. Now it says Verified, Unverified or Not Checked, and Fact Check counts the claims it actually supported.
  • A right answer doesn’t get “Verification failed” under it. One check compares the whole answer against a single passage, and it was failing answers whose every sentence the claim check had already supported. When that’s the only check that fails, the claim check decides.

Private Cloud Compute

  • “Just Once” means once. In Deep Think and Maximum, choosing Just Once on the consent sheet carried over, so the next Deep Think or Maximum question could go to Private Cloud Compute without asking. Now Just Once covers one question.

Getting started

  • It tells you when your device can’t run Apple Intelligence. Chat used to say “Enable it in Settings” even on a device that can’t turn Apple Intelligence on, and Settings said “Preparing AI Models…” forever. Now the app tells you which (not supported on this device, turned off, or still downloading) and what still works without it.
  • An empty library doesn’t suggest questions it can’t answer. With no documents in a library, the chat suggested four prompts and every one of them came back “No documents yet.” An empty library doesn’t suggest anything now.
  • The hardware readout waits for the welcome screen. On a first launch it sat on top of the welcome screen. Now it shows up once you’re in the app.

v5.5 - September 30, 2026

One fix and lower prices this time. A “how much” question could get answered with a number from a totally different document, with a Verified badge on it.

Answers

  • “How much” questions stop grabbing numbers from the wrong document. On my Mac, “How much notice do I have to give before I move out?” came back “1 lb.” from an air fryer manual’s cooking chart, marked Verified. Any “how much” or “how many” made the app treat every measurement it found as a possible answer, and a weight like “1 lb” could pass for a liquid. Now it only goes looking for a volume when you ask about fuel, gas, oil, coolant or a tank, or what something holds, so a question like the notice one goes to the model instead. The model still got that one wrong on my Mac (it said at least 30 days, the lease says 60) and still marked it Verified, so I’m not done with it.
  • Capacities like “4.5L” count as liters. When you ask how much something holds, “4.5L” counts the same as “4.5 L” now.

Plans

  • Pro costs less from October 1. Pro is $4.99 a month or $24.99 a year in the US, down from $5.99 and $29.99, and it’s lower in most other countries too. If you’re already subscribed, you renew at the lower price automatically. Lifetime is $49.99 now, down from its $59.99 regular price, and it’s just called Lifetime.
  • Each plan does the math for you. The plans screen shows what Annual saves over twelve months of Monthly and what it works out to per month, and it says when Lifetime costs less than a year of Monthly. Every number comes from your own App Store’s prices, so it’s right in your currency.

v5.4 - September 24, 2026

Answers

  • Finished answers don’t sit there anymore. If you asked Standard for one specific fact, it ran a second check after the answer was already written, and the cursor kept blinking with the text box locked (19 seconds to almost three minutes on my Mac). That check didn’t change a single answer I measured, so now it only runs when the app’s own verification flags something.
  • You can type while an answer gets checked. When the text is done, the cursor goes and the answer says it’s checking sources. Send your next question or hit Stop and the answer stays, sources and all.
  • Deep Think and Maximum write in front of you. They used to show nothing until every single step was done, then dump the whole answer at once. Now you watch the final answer get written for most questions, with “Refining…” while a later step might still improve it.
  • Short answers stream too. An answer small enough for the app’s structured format used to get written in one piece and then played back. Now it shows up as it’s written.

Background

  • You’ll know when an answer’s done. On iPhone you get a light tap when it finishes, plus a badge on the Chat tab if you were on another tab. The timer keeps running when you come back to an answer that’s still going, instead of freezing at the moment you left.
  • An answer cut off in the background says so. If iOS ends the app’s background time before an answer finishes, whatever it already wrote stays in the chat with a note saying why. It used to just disappear.

v5.3 - September 18, 2026

Plans

  • The upgrade screen leads with what a plan actually changes. It used to list storage numbers first, “up to 1,000 documents”, “10 libraries”, and said nothing about the model. Every plan runs the same model: Apple Intelligence on your device, Deep Think, and Private Cloud Compute after you approve what would be sent. What Pro and Lifetime change is the daily cap on Maximum mode, three runs a day on the free plan and none on a paid one, plus how many documents and libraries you can keep. The screen now says that, in that order, and the sample Product Guide’s plan section matches.
  • When Lifetime is on sale, the app says so. One line at the top of the Chat and Documents tabs, for anyone who has not bought, with the real percentage and the real last day. Both numbers come from the App Store’s own price for your country against the regular price, so the line cannot claim a discount that is not running. Dismiss it and it stays away until a new sale.
  • A thumbs-up now offers to write a review. Apple’s rating sheet gives stars and never a written review, and a written review is what someone reads on the store page. So when you thumb-up an answer, the app asks once per version whether you’d rate it, write a review, or send feedback. And the Lifetime card now says how many months of Pro Annual its price buys, from the store’s own prices.
  • One look at the plans, once. After the fourth answer the app has verified against your files, one answer after it asks for a rating, it shows you the plans screen. It says it will not ask again, and it does not. If you have already bought, you never see it.
  • Pro Annual no longer offers a free trial. It is one payment a year. If you started a trial before this change, nothing about it changes.

Getting Started

  • The sample library’s suggested questions are back. They were written by hand for the three sample documents, and one leftover duplicate of a sample (“RAG Technical Architecture 2”, left behind by an earlier refresh) was enough to switch them off: the app decided the library was no longer the sample library and built questions from templates instead, which is where “What is nothing?” and “What is the silicon?” came from. Three fixes. Duplicate sample copies are removed on the next visit to the Documents tab, and only when the original is also there. The hand-written set shows whenever the library holds the three samples, duplicates or not. And a library whose questions were built before Apple Intelligence was ready, which is exactly when the sample library is created, gets them rebuilt by the model the next time it is available, instead of keeping the template set forever.
  • No suggestion is better than a bad one. When the model is available and has nothing well-grounded to ask about a document, the app now shows nothing for it rather than a template question. The templates also stopped treating a lone word like “nothing”, a phrase from the document’s own title, or any passage containing the word “study” as material for a question.
  • The welcome screens follow your light or dark setting now. They were always dark, whatever the rest of the app was doing, so turning the phone to light mode gave you a white app with one navy screen at the front of it. In light mode the status bar was the real casualty: iOS draws the clock and battery in dark ink on a light screen, and the welcome screen underneath was nearly black, so they simply vanished.
  • Light is its own design, not the dark one flipped. Inverting would have given a flat pale rectangle. The light version is a warm off-white settling into a faint blue-grey, with the glow behind it eased back, so it still feels like a lit surface rather than a blank page.

Reading your documents

  • The app now uses Apple’s own detector to find the real things inside a document, instead of its own pattern matching. When a page is read, addresses, phone numbers, dates, amounts of money, measurements, flight numbers and tracking numbers are recognised for what they are rather than matched as text that looks roughly right. The patterns it used before only understood US phone numbers and three currencies, so a European invoice produced no amounts at all and a postal address was never recognised. An amount is now recorded together with the currency it is in, and a date in a standard form, rather than as whatever characters happened to sit on the page, which is what makes them findable later.
  • Known limit, stated plainly: this currently applies to text inside tables. Prose on the same page does not yet contribute.

Answering

  • Three sliders in Model Parameters were telling you they did something they cannot do. Frequency Penalty, Presence Penalty and Repetition Penalty sat under a line saying they affect how the model picks words. Apple’s generation API has no penalty setting of any kind, so Apple Intelligence ignores all three, on this device and on Private Cloud Compute. They now say so. They are still there, and still adjustable, because they are for a local model host on Mac that is not built yet.
  • New, and off until you turn it on: Adapt to the question. In Model Parameters, this lets the app choose how careful or how expansive an answer should be from what you actually asked. Looking up a value or a figure becomes near-repeatable, so asking the same question twice gives the same number. Comparing two things, or investigating a topic, gets more room to work. Your own saved settings are not changed, and turning it off restores them immediately.
  • It works in Deep Think and Maximum now too. When it first shipped it only applied to Standard, which was most of the point missed: those are the modes where how carefully a question gets answered matters most. They use the temperature it picks and keep their own per-step lengths, because those run many passes with budgets chosen for what each pass is doing.
  • You can write your own reasoning profile. Apple’s control takes three fixed settings or a description you write yourself, and the app had only ever used the fixed three. In Model Parameters you can now write something like “work backwards from the conclusion and say which step is weakest” and it replaces the setting for questions that reason. This is separate from the app’s own reasoning: Deep Think and Maximum think across many steps on your device and that is unaffected by what you write here. What this changes is how hard Apple’s model works inside a single answer, which Apple only allows on Private Cloud Compute.
  • Why it is off by default: nobody has shown it produces better answers yet. It is a setting you can judge for yourself, not a change made on your behalf.

v5.2 - September 10, 2026

The release that turns on Private Cloud Compute. Nothing about how the app reads, searches or checks your files changes; what changes is that one step, the writing of the answer, can now leave the device when it needs to, and only after you say so. Requires iOS, iPadOS or macOS 27.

Plans

  • When a launch discount is running, the plans screen now shows what the price normally is, how much is off, and the day it ends. The App Store does not mark a reduced in-app purchase as reduced, so without this the price simply looks lower and there is no way to tell it is temporary. The app only says a discount is on when the price it is being charged is genuinely below the normal one, so it cannot advertise a sale that is not running.

Private Cloud Compute

  • When a question needs more room than the model on your device can hold, the writing step can now go to Apple’s Private Cloud Compute. Reading your files, searching them, choosing what to cite and checking the finished answer against the passages still happen on your device, exactly as before. Every release since 4.6 has carried this support compiled out, waiting on iOS and macOS 27. This is the first build made with the toolchain that compiles it in.
  • Nothing leaves without asking. Before anything is sent, a sheet shows what would go: how many passages, how large, and why. Allow it once, allow it always, or keep everything on the device. You can also pin routing to On-Device in Settings and the question never comes up.
  • Every answer says where it was written. Expand the metrics bar under an answer to see On-Device or Private Cloud Compute.
  • Apple’s servers keep nothing after the answer. The connection is end-to-end encrypted and the request is not accessible to Apple or to the developer; these are Apple’s stated properties of Private Cloud Compute, and the app adds nothing to the request beyond the passages you approved.
  • Measured on an iPhone with the A18 Pro: roughly 86 tokens per second from Private Cloud Compute, against 27 on the device. Those figures were taken on a local build with the iOS 27 toolchain and are the same figures the earlier release notes cited.

Honesty

  • How It Works and the About screen said the app asks before sending to Private Cloud Compute, on builds that could not send anything. The check they read answered a different question, whether the operating system has Apple Intelligence, which is true of every iPhone that can run the app. It now answers the real one: whether this build can route to Private Cloud Compute. On a system older than 27 those screens say so plainly.
  • The built-in guides, the Glossary and the Settings capability list now describe the build you are running. Each of them picks its wording on the same condition the code uses, so a build with Private Cloud Compute says it is on and a build without it says it is not.

Ratings

  • The app now asks for an App Store rating, once, after your third verified answer. It never asks on launch, never after an answer it could not verify or declined to give, and never twice within four months; Apple’s own sheet decides whether to show at all and caps it at three times a year. If you would rather not be asked, dismiss it once and it stays away.

v5.1 - September 2, 2026

Most of this release is about the Mac, where importing a large document had become slow enough to be unusable. The text recognition changes apply to iPhone and iPad as well.

Built-in Guides

  • The three sample documents now describe the app you are running. They said complex questions were sent to Private Cloud Compute, which nothing in this build does; they listed fewer formats than the app reads; they described the quality modes by where they run rather than by what they change; and they said nothing about importing, which is where most of the recent work went. All three are rewritten from the code: every format, what happens during an import and after a quit, what each quality mode actually changes, the plan limits, the measured speeds, and what Private Cloud Compute will add when it arrives with iOS and macOS 27. If you already imported them, they update in place.
  • The What’s New screen has entries for 5.0.2 and 5.1. Updating to either used to show nothing, because neither release had been written for it.

Appearance

  • The Mac app icon was a square, and its dark version had a blue rim. Every Mac icon slot was a solid square filling the whole tile, with no transparency at the corners. Mac icons are supposed to sit inside a rounded shape with a margin around it, which is why yours looked like a hard tile in the Dock next to everything else. The dark version had a second problem: it had been made by repainting the background behind the bulb without redrawing the edges, so a thin line of the original blue was still there along every outline. Both are redrawn. The iPhone and iPad icons were already correct and are unchanged.

Speed

  • Every PDF page was drawn at four times the resolution it asked for, then copied through an uncompressed image format and back. The Mac used a drawing method Apple deprecated for exactly this reason: it sizes its canvas from the display rather than from the request, so on a Retina screen a page asked for at 3060x3960 was actually drawn at 6120x7920. That result was then encoded to an uncompressed image in memory and decoded straight back, once per page. Measured on this hardware, roughly 370 MB written and read per page, for nothing. The iPhone never did this; its half of the code was already correct. Each page is now drawn once, at the size requested. The wall-clock effect on a full document has not been measured and no speed multiple is claimed here.
  • With the app open and doing nothing, it was re-reading every library’s search index about seventeen times a second. A background timer kept finding one library in a state it could not resolve, and each pass reloaded every index whether or not anything had changed. Over a measured 164 idle seconds that was 2,848 index reads, the same index re-read 288 times. Left running for about three hours it accumulated a backlog of queued work it could never finish and the app stopped responding. Indexes are now re-read only when they have actually changed on disk.

Reliability

  • A large import paused by quitting the app looked like it had lost everything. Reopening showed the document sitting at zero. The work was never lost: the app has always resumed from the last page it committed. But the screen said otherwise, and removing the item is the one action that discards that progress, so the display was steering people toward the only destructive choice available. A resumed import now reports how many pages it already finished.

Text Recognition

  • The app asked the system to guess each page’s language while also handing it a list of thirteen to choose from. Those two instructions contradict each other. Apple’s own guidance is to state the languages when you know them, because the automatic guess is not guaranteed correct, and a wrong guess means text is corrected against the wrong dictionary, which damages it rather than merely slowing it down. The language is now determined once per document from the text the file already contains, and only that language is used.
  • Page images went to text recognition at the highest possible resolution with no downscaling. That is the slowest and most memory-hungry setting the system offers, and it was set deliberately with a comment saying it caught all text sizes. Recognition now scales down to what is needed to read roughly three-point type, which is smaller than the fine print in any real document.

v5.0.2 - August 27, 2026

Mac only. On the Mac there was no way to get a document into the app at all: the Add Documents button opened nothing, and dragging a file onto the window did nothing either. Both are fixed. iPhone and iPad are not affected by this release and are unchanged.

Getting Documents In

  • The Add Documents button opened nothing. The button asked macOS for a file picker at the one moment the system refuses to open one, so the request was discarded and no window ever appeared. The same fault hit the two file buttons inside a chat, where it was worse: those open into a panel that is otherwise empty, so there was no button to fall back to and no sign anything had gone wrong. All three now ask at a point the system accepts.
  • Library Settings was unreadable on the Mac. The screen was being drawn as a narrow strip down the left with a large blank area beside it, squeezed hard enough that words broke apart mid-way — “Documents” came out as “Doc ume nts”. It was built on an older navigation container that macOS turns into a two-pane layout meant for a sidebar and a detail view, which is not what a settings screen wants. It now uses a single column at a sensible width. Other screens still use that older container and may show the same thing; those are being worked through separately.
  • The built-in sample documents were quietly duplicating themselves. When a sample is corrected in a new version, the app replaces your copy with the updated one. It was deleting the original but not any duplicate an earlier update had already left behind, so each round added another. One library ended up with five documents for three samples. This matters beyond tidiness: the app answers out of these documents, so a duplicate meant the same passage counted twice when it decided what to quote. Existing duplicates are cleaned up on the next update, and documents you named yourself are never touched, even if the name looks similar.
  • You can now drag files from Finder straight into a library. This had never worked, because nothing in the app was listening for a dropped file. The whole library area now accepts them, so a drop does not have to be aimed at anything precise, and dropped files go through the same size and quota checks and the same import review as files picked with the button. Folders are not accepted yet, so drop the files from inside them instead.

v5.0.1 - August 26, 2026

macOS 5.0 went to the App Store as an earlier build and is missing everything below. This release brings it level. On iPhone and iPad this is the first 5.x release, so the v5.0 notes below are part of what you are getting.

Speed

  • The Documents tab stopped waiting on a number it never showed you. Opening the tab first went and counted your cached documents, and only then got on with appearing. That count feeds one optional row that is hidden unless the count is above zero — and on the traced device it was zero, so the row was never drawn. The count is now fetched quietly in the background while the tab opens immediately. On the device this was measured on it cost between 29 and 393 milliseconds every single time the tab was opened, and it was the slowest thing in a log of nearly six thousand lines. Nothing about your documents, your libraries or search changed; the tab just stops waiting to be told something it does not display.

Your Hardware

  • Rotating the device left black rectangles around the floating hardware readout. The readout lives in its own small window that positions itself, and nothing was repositioning it after a rotation. It was also being held inside a portrait-shaped area even in landscape. Both fixed. Dragging it is unchanged, and the chip outlines still stay pinned to where the hardware physically is, because the hardware does not move when you turn the device.
  • The Temperature slider did nothing on one of the three sampling settings, and did not say so. Predictable always takes the most likely next word, so temperature cannot change its output. The slider stayed adjustable anyway. It is now dimmed on that setting, with a note about where it does apply.
  • The Atlas was labelling a medical paper “API Reference”. Cluster names were matched on fragments rather than whole words, so “api” was found inside “therapies” and “term” inside “determined”. A paper about dopamine and serotonin was filed under categories written for software documentation. The same list existed in three places and all three were wrong.
  • “Analyzing corpus…” was never analyzing anything. Every library that was not the one you had open showed a spinner that could not finish, because nothing had been asked to run. Reading a library’s settings means opening it and loading every passage, so doing it for all of them at once would make that screen slow instead. It now tells you it has not run yet, and gives you a button.
  • The library selector on the Database screen is a row, not a menu. With eight libraries, a menu meant a tap to see the options and a second tap to choose one.
  • A tag that appears once in a document no longer describes it. Running headers and page furniture were being turned into tags, which is where “pychatry” came from — a mangled journal masthead read off the top of a page. A word that genuinely describes a document appears in it more than once.
  • Turning performance up was making the app slower at the thing it does most. The four execution profiles set which parts of the chip do the work. Choosing Maximum switched the Neural Engine off and left it doing less than Performance, and when building your search index both Performance and Maximum switched it off, so the middle setting was the only one using the whole chip. Every setting now uses everything the one below it used, and more.
  • The performance selector shows you what it is doing. It was a dropdown that displayed one option and hid the other three. It is now four cards side by side, each showing which parts of the chip it uses, with a plain description of exactly what changes when you pick one.
  • Deep Think never ran a minimum number of passes, and the screen said it did. It said “6 to 8 reasoning sessions”. There is no minimum; it stops as soon as your documents have nothing new to add, or once it is confident enough. One recorded run stopped at five. It now says “up to 8” and explains both reasons it stops early.
  • Macs were being given iPhone-sized limits. The app sorts devices into performance tiers, and the tier below the top one is tuned for phones. Every Mac except the Pro, Max and Ultra models had been placed in it, so a MacBook with 32 GB of memory was processing 1,024 items at a time when it can handle 6,144, and was not scaling with its memory at all. Every Apple chip now sits in the tier its actual capability warrants. Heat is still managed by measuring the machine’s real temperature while it works, which is what was protecting it all along.
  • The hardware panel now shows free memory, and one of its numbers was nonsense. The small floating readout in the corner listed processor, Neural Engine and graphics activity but never how much memory was left, which is the number that decides whether a large import will finish. It does now. Separately, the Hardware Envelope card claimed a limit of 1,073,741,824 simultaneous threads; the real figure is 1,024, and the wrong one came from multiplying three separate limits together instead of reading one.
  • The app understands Apple chips that do not exist yet. Chip support was written out by hand and stopped at the M5, so a Mac newer than that was labelled generically and given settings tuned for an M3. It now works out the generation from the chip itself and scales to match. Four other cases where an unrecognised device quietly switched off four parts of search have been closed as well; the only devices still treated as unsupported are ones that genuinely cannot run Apple Intelligence.

v5.0 - August 10, 2026

Documents were quietly losing parts of themselves, answers were built from a fraction of what was found, and the app was rewriting your library on every launch. This release is the search for all three.

Speed

  • Deep Think answers roughly three times faster. The same question that took four and a half minutes now takes eighty seconds, and comes back slightly longer. It was re-reading the same passages three times over on smaller libraries; it now stops once it has covered the material.
  • The app starts faster. A 43 MB model was being loaded every single launch, before anything appeared on screen, whether or not you went on to ask a question or import anything. It now loads the first time it is actually needed. If you open the app to check something and close it again, that work never happens.
  • Deep Think stops when it runs out of new material instead of re-reading. On a smaller library it would run its full eight reasoning passes even after it had already read everything, repeating the first three passes word for word. Those repeats took about a third of the total time and could not tell it anything new. It now stops once it has covered the material. It still reads every part of every document it retrieved, and it still keeps going when there is genuinely more to read.
  • A document you import no longer loses its searchability to the app’s own housekeeping. The part that makes a document searchable is written as soon as it finishes processing, and the record saying the document belongs to the library is written a little after that. In between, the background job that keeps your libraries consistent across devices looked at a library that appeared to contain nothing, concluded its search index must be junk, and deleted it. The document stayed in the list. Its searchability was gone, and nothing told you. That job now refuses to delete an index that still holds content, and every deletion it does make says why it made it.
  • Switching libraries during an import no longer splits the document from its index. If you started importing a file and then moved to another library while it processed, the file was filed in the library you started in but the part that makes it searchable was written to the one you moved to. The document showed up in the list and in the database view, and then answered nothing, which is the state that made a library look broken for no reason. An import now always writes to the library it started in, however long it takes and wherever you go in the meantime.
  • A library that has lost its ability to answer now tells you, and can repair itself. A library could keep its file list and its text search while the part that actually answers questions quietly went missing, and detection previously ran only when you switched libraries, so it could sit broken for an entire session with nothing to tell you. The check now also runs the moment a question comes back with nothing, and the repair is confirmed working on a real device: detected, rebuilt and answering again in about four seconds. Two related cases, where the repair is blocked because files are still importing, are fixed in code but have not yet been reproduced on a device to prove it, so they are described here as fixed rather than as verified.
  • The app stopped re-reading your entire library to decide it had nothing to do. Keeping libraries consistent across devices means comparing what is on this device against what is in iCloud, and the comparison was reading every passage of every library out of both places before it could decide anything. On one device that came to 43,000 passage records read in a single session, half of them before the app had drawn a screen, and it wrote nothing at all. It now checks whether anything has actually changed before reading, and only does the full comparison when something has. The comparison itself is unchanged, because it is the thing that stops good data being overwritten.
  • iCloud sync stopped re-uploading libraries that had not changed. Every sync pass rewrote each library’s search index whether or not anything was different, and because rewriting a file looks like a change, that started another pass. One launch could do this six times over, writing hundreds of megabytes and queueing half of it to iCloud, all of it identical to what was already there. Sync now compares before it writes, and skips when nothing differs. Nothing about what gets synced has changed, only how often it is rewritten.
  • The source chips under an answer scroll properly now. They used a custom press handler that fought the sideways scroll, so swiping across them sometimes registered as a tap instead. They look and respond exactly the same, they just no longer argue with the scroll.
  • Leaving the chat tab no longer kills the answer you were waiting for. Switching to Documents mid-answer and coming back used to cancel it and discard everything written so far, with nothing to tell you why. The answer now keeps going while you look at something else.
  • The chat no longer loses your place. Scrolling up to re-read an older answer and switching tabs used to slam you back to the newest message on return. It now stays where you left it, and still follows along automatically when you are reading the latest.
  • Atlas keeps showing what it already worked out. Re-opening the tab replaced the whole page with a loading spinner while it recalculated the same result. It now leaves the existing view up until the new one is ready.
  • Answers stream more smoothly. While an answer was arriving, the app re-formatted the entire text from scratch up to fifty times a second, on the same thread that draws the screen. Formatting now happens once, when the answer finishes. The finished answer looks exactly the same, and text stops flickering as half-finished bold and code blocks resolve.
  • Opening the Atlas tab did its work twice. Two separate triggers were both starting the same analysis. Now one does.

Libraries

  • Every action above your documents is now one tap. Add, search, library settings, emptying a library and deleting one were two buttons and a three-dot menu hiding the rest. They are now five icons on a single row, nothing hidden. VoiceOver still reads each one’s full name, including which library a delete would affect, and both delete actions still ask you to confirm and still tell you exactly what will go.
  • Press and hold on a library now behaves like the rest of iOS. It was using a custom gesture that competed with sideways scrolling, so holding a library sometimes scrolled the row instead and nothing told you which one was about to happen. It now uses the standard press-and-hold menu, with the usual preview and haptic.
  • Creating a library no longer suggests a name you already have. The suggested name was based on how many libraries you had rather than what they were called, so after deleting one, the next suggestion could collide with a library still on screen. Accepting it left you with two libraries sharing a name.
  • Importing certain documents could freeze the whole import queue. One kind of file-path lookup could deadlock partway through, and because everything after it in the queue waits its turn, nothing else would import either until the app was force-quit. Found from a live stack sample and fixed at the root.
  • Switching between libraries no longer resets your place or flashes the screen. The library picker at the top of Documents was rebuilt from scratch every time you switched to or from an empty library, which reset its scroll position back to your first library and made the whole screen visibly redraw. Switching libraries now leaves the picker exactly where it was.

Your Documents

  • A document’s auto-generated tags no longer come from its bibliography. The app samples the start, middle and end of a document to work out what it is about, and on a research paper the end is the reference list — so a third of what it read was author surnames. The same paper produced “analysis, depression, dopamine, motivation, serotonin” on one import and “cho, merten, pychatry, zeng” on another, the second set being cited authors and a mangled page header. Reference sections are left out of that sample now, and a capture on August 26 confirms it working: 137 reference-list passages were excluded from the sample on that same paper. The mangled header half of this is not fixed and is still visible. In that same capture the tag “pychatry” appeared again, and it does not come from the bibliography at all — the journal prints its name sideways down the edge of every page, the scanner reads that sideways strip badly, and the result is treated as a heading. That is a separate defect in how page furniture is read, it is recorded, and it is not claimed as fixed here.
  • Tables in Word documents were being thrown away. Each one was read into rows and then dropped, so a document could import looking fine with all of its numbers missing.
  • Images keep their layout. Every image became one unbroken line of text before anything downstream could read it.
  • Photographing a page now matches importing it. Camera captures came out as flat text where the same page imported as table cells.
  • Scanned pages report their scanning honestly. A fully scanned PDF used to tell you it had scanned zero pages.
  • Text from a PDF page is no longer shuffled into the wrong order. Lines were put in sequence by a comparison that treated any two lines closer together than a fixed distance as being level with each other, and then ordered those by their left edge. That distance was larger than the gap between lines in a typical journal article, so almost every pair of neighbouring lines was ordered by indentation instead of by position down the page, and came out swapped. Every word had been read correctly; a paragraph simply arrived as a jumble of half-sentences, which is why nothing flagged it. Lines now sort by position down the page, and the left edge only settles an exact tie.
  • A page the parser knew it had read badly is no longer repaired and then discarded.
  • PDFs with figures but no tables keep their figures.
  • Pages, Numbers and Keynote files now fail clearly. They were never actually readable; importing one no longer looks like it worked.
  • More than half of every document never reached the part that makes it searchable. A padding setting in the two bundled tokenizer files capped text at 128 tokens against models built for 512, and separately made the length check return the same number for every input, so nothing downstream could notice. Measured with the real vocabulary over 139 live chunks: 90% of chunks were cut short and 55% of all library content never reached the embedder. The cap is gone and the length check counts again.
  • The model that reads meaning was looking at one position instead of the whole passage. all-MiniLM-L6-v2 is trained to average across a passage, and the app’s export read a single position the model was never trained to carry the whole meaning. The vectors were never garbage, which is exactly why nothing caught it: correctly sized, correctly normalised, and encoding far less than they should. Re-exported to average properly, this moved rank-1 retrieval from 0.000 to 0.571 across 21 paired test questions, 12 better and none worse.
  • Libraries you already have will now offer to rebuild their search index, once. This release changed how your text becomes searchable, twice over: more than half of every document was being cut off before it was ever indexed, and the part that reads meaning was looking at one position in a passage instead of the whole thing. Anything indexed before this update was built the old way. The app could not tell the difference before and would have left those libraries quietly worse forever. It now notices and offers a rebuild, which you can take when it suits you — nothing is deleted, and the library keeps working in the meantime.

Answers

  • A long answer is no longer replaced by a one-line stub. After the app writes a long answer, a final stage re-reads every claim in it against your documents. That stage was given a smaller budget than the answer it was checking, so on a long answer it could be shown as little as one and a half percent of it and then replace the whole thing with a fragment. It now keeps its share of the budget for the answer rather than letting the evidence take all of it, and if it still cannot see enough of the answer to judge it fairly, it declines and leaves the original alone.
  • The app stopped treating research papers as parts catalogues. A question about dopamine was being classified as a specification lookup, because the list of words that signal “this person wants a spec value” contained “min” and the app was matching it anywhere in the text rather than as a whole word. Dopa-min-e. That switched on scoring built for datasheets, which promotes tables and lists over prose, and in a journal article the tables and lists are the reference section. The same fault made “often” and “after” look like feet, and “example” look like amps. Reference markers such as “behavior.42” were being read as part numbers.
  • Reference lists no longer get cited as sources. A bibliography is the worst possible evidence and one of the best keyword matches: a page of citations about dopamine says “dopamine” many times and states no fact about it. Two of twenty sources in a real answer were citation entries. The app now recognises the shape of a reference list and ranks it below prose, while leaving a genuine question about what a paper cited alone.
  • Search ranks section headings again. A weighting mistake had dropped section paths out of ranking entirely.
  • Evidence that retrieval had already found is no longer discarded when sentence extraction matches nothing.
  • Long questions can now reach Private Cloud Compute. They were being kept on device precisely when they were too big for it.
  • The evidence behind an answer was being put in an undefined order. After the app ranks passages, one more pass re-orders them so that a passage with a clear advantage can outrank a merely similar one. The rule it used compared two passages as equal when their scores were close, and as different when they were far apart — which sounds reasonable and is mathematically inconsistent, because three passages can each be close to the next while the outer two are far apart. Given a rule like that, the sort has no defined answer, and the evidence order came out effectively arbitrary. It now uses a rule that always agrees with itself.
  • Deep Think was writing its answer from a fraction of what it actually found. It could retrieve the right passages and then discard most of them before writing a single word of the answer, including the single best one it found. It now keeps what it retrieved and, in the technical detail view, says exactly what it chose to leave out and why.
  • An answer that cites your documents can no longer be replaced by one that cites nothing. After writing an answer, the app sometimes tries to improve it, and a rule exists to stop a polished-but-unsourced rewrite from displacing a properly cited one. Two holes in that rule are closed. One editing pass was allowed to replace the answer on length alone, without the citation check — and that pass is the one most likely to strip citations, because it is explicitly told to remove unsupported claims. Separately, the citation counter only recognised sources written as [1] and not as (1), so an answer that cited everything in the second style counted as citing nothing, and the protection switched itself off.
  • An answer can no longer tell you your documents are silent while quoting them. The check meant to catch a bad answer was only confirming that its citation numbers pointed at real passages, which they did — so an answer that said “the documents contain no evidence of this” while citing twenty of them was certified at 88% confidence. That specific contradiction is now caught and the answer is regenerated. An honest “I couldn’t find this in your documents”, with nothing cited, is untouched — that one is true, and it stays.
  • Citations now always point at a source that’s actually there. An answer could cite a source number past the end of its own source list, and nothing caught it. Citations are checked against the real list before you see them, and the confidence score can now report a genuine failure instead of always settling on a reassuring middle number.
  • Combining keyword and meaning-based search stopped losing to keyword search alone. The two search methods are supposed to complement each other, and instead the combined result was ranking worse than keyword search by itself. The keyword search’s best matches can no longer be crowded out before the two are compared.

Your Sample Documents

  • The three sample documents described things the app cannot do. They claimed Pages, Numbers and Keynote support, credited the wrong framework, and documented a screen that does not exist. The app answers questions out of those documents, so a wrong sentence became a wrong answer.
  • If you already had them, they update themselves once. You will see a short re-import the first time you open Documents, and a notice explaining it. Nothing you imported yourself is touched, and samples you deleted stay deleted.

Settings

  • Settings is a searchable list instead of one long scroll. Fifteen stacked panels became about ten rows in sections. Nothing was removed. Type “temperature” and you land on it.
  • The generation controls are reachable. Temperature and response length were built with no way into them. Settings → Advanced, one tap. Both work on device and on Private Cloud Compute.
  • You can choose how the model picks its words. Top-K, Top-P or Greedy. The app used to decide for you and always picked the same one, so the Top-P slider did nothing. Unchanged unless you change it.
  • Answers can be made reproducible. Turn on “Reproducible answers” and the same question against the same library returns the same answer every time. Useful if you are checking work, or comparing two libraries fairly.
  • Controls that never affected Apple Intelligence now say so. The three penalty sliders apply to a self-hosted model, not to on-device or Private Cloud Compute answers.
  • Five switches that did nothing are no longer switches. Writing Tools, Speech Analysis, Translation, Screen Awareness and Image Playground are real and always on. The toggles never controlled them.

Every Word, Explained Where You Read It

  • Tap any figure the app shows you and it tells you what it means. “38 TOPS”, “32/batch”, “768 search”, “Chunks”, “Vectors”. Where they already are, without leaving the screen.
  • Two explanations for each, not one. A plain one first, and the mechanism underneath if you want it. Turn the technical version on once and it stays on everywhere.
  • The four import stages explain themselves while they run. Extract, Chunk, Embed and Index are all tappable during the import you watch on first launch.
  • Nothing is buried. Settings has the full list under Plain English, searchable, including by the technical name if that happens to be the word you know.
  • Tapping a Plain English term now reliably opens its definition. It could animate as if it were opening and then leave you looking at the same list, and backing out could land you on a blank screen. Terms now open the same way every other definition in the app already did.
  • Plain English covers a lot more of the app now. Seven new entries: the difference between Standard, Deep Think and Maximum, where an answer actually gets written, and the trust cluster — confidence, fidelity, the live verification checklist, and what the red “Abstained” badge means. Confidence and fidelity in particular are two different measurements shown together with nothing previously explaining that they can disagree, or why.

Your Libraries

  • “Remove Local Copies” is now “Remove All Documents”, because that is what it did. It never only removed local copies. It deleted those documents from iCloud and from your other devices too, while telling you Sync Now could bring them back. It could not. The wording now says what happens, and names the library it applies to.
  • Deleting one document said the same untrue thing, and its button is now just “Delete”.
  • Emptying a library used to leave things behind. The documents stayed in Spotlight search, and their files stayed on disk taking up space. Both are cleaned up now.
  • Moving a library off iCloud asks first. One tap used to remove that library’s copy from iCloud with no warning at all.
  • The Documents toolbar no longer hides half its buttons off the edge of the screen. Library Settings, and the two destructive actions, now live in one menu you can see. Visualize is gone from that row, because it only switched you to the Atlas tab that is already at the bottom of the screen.
  • Two libraries with the same name can be told apart by a short code on the chip.
  • Library chips in Semantic Search no longer show storage buttons that did nothing.

Your Database Tab

  • You can look at any library from here. It used to offer only “All Libraries” or whichever library happened to be active, so seeing a different one meant going to Documents, switching, and coming back. Every library is now in one menu.
  • “All Libraries” actually shows all of them. The document list underneath used to show only the active library no matter what you picked, and never said which library you were looking at. It says now.

Library Settings

  • Changing the embedding model no longer wipes your vectors before you agree to rebuild them. It deleted them at Save, then asked, and choosing “Later” left the library holding documents it could not search.

  • The chunk size sliders now stop where the app actually stops. They went up to 600 words and 200 overlap while importing quietly capped them at 260 and 50, so over half of each slider did nothing. If you had set 400, you were already getting 260.

  • The chunk size controls now say you should not touch them. The app already picks a chunk size from the kind of file you imported, smaller for code, larger for reports and transcripts. Setting these by hand replaces that with one size for everything in the library.

  • Two descriptions on that screen were not true. One said chunks are split by comparing meaning between sentences; that never runs. It splits on section headings and a fixed list of ten English phrases, and the screen says so now, including that the list is English only. The other described the storage format the app stopped using.

  • Deleting a library from its settings screen no longer half-works. If iCloud refused the delete, that screen used to remove the library from this device anyway, so the next sync brought it back and you were left thinking it was gone. It now stops, keeps the library, and tells you why.

  • The two destructive actions no longer look like the same button. They sat next to each other, both red, both a bin icon, and one empties a library while the other removes it. They are now “Remove All Documents from X” and “Delete the X Library”, with different icons and a divider between them.

  • “Cached Documents” is hidden until there is something in it. It was always empty, because the feature that would fill it has not been built yet.

  • On a Mac, a Shortcut running in the background could send a question to Apple’s Private Cloud Compute with nobody there to approve it. The check that requires you to be looking at the app was only ever running on iPhone and iPad. It runs on Mac now.

  • A question is no longer sent to the cloud when the app cannot tell how much cloud quota is left. It now answers on this device instead of guessing.

Things The App Was Claiming That Weren’t True

Settings describes what the app is doing while it answers you. Several of those lines described things the code does not do, so they are gone or corrected.

  • Three places in the app still described Private Cloud Compute as something it does today. The App Store text and the README were corrected in August; the copy inside the app was not. The Model Info card advertised an “On-device model + PCC server” because it checked whether Apple Intelligence was available rather than whether PCC was. How It Works said the app “asks whether to send that single request” to PCC, which it never does in a shipping build. And the sample documents the app reads back to you as fact carried a section headed “When does PCC activate?” beginning “The app routes to PCC automatically”. All three now describe the build you are actually running, and will describe PCC in the present tense again on a build that contains it. The measured speed figure was kept and labelled with the build it came from rather than deleted, because it is real.
  • Settings listed eight agentic tools, and all eight were the wrong ones. Four tools are actually wired up. Those four are now what it names.
  • Two advertised features did not exist, and are no longer advertised.
  • The model picker listed a tier the app cannot select.
  • Speed figures that were never measured have been removed, including from the sample documents the app reads back to you as fact. Where the underlying work was real, it is now described by what it does instead of by a number.
  • Every remaining capability line was checked against the code that would have to run it.

The App Itself

  • The upload status stops breaking its own words in half. With the controls beside it taking the room, “Processing uploads” could render as “Process / ing / uploads”. It is the one piece of chrome on screen during every import.
  • The app icon now follows your device’s dark mode. The existing light mark stays the same, and a matching dark appearance is selected when your device uses dark mode.
  • Your device is identified correctly. iPhone 17 and M5 hardware reported itself as “A12 or Older” with limited performance.
  • The first screen no longer cuts off its own text. Half the headline and all three example questions were being truncated.
  • The chat controls match each other, and the mode menu explains itself. It also shows which mode is active, and how many Maximum runs you have left before you pick it.
  • The hardware readout follows a live answer and no longer stays lit after you stop one.
  • The screen readers can reach the hardware panel properly. Its figures used to be read as one long block; each one is now its own element with its own definition.
  • Answers that ended in bold or italic text no longer come out broken. The last two characters were being cut off, which left the formatting unclosed and bled it into the rest of the screen. The clearest case was the notice you get when an answer could not be verified: its closing bracket was always missing.
  • The embedding view no longer claims your vectors have 512 numbers when they have 384. It reads the real number from the library you are looking at, which differs depending on which embedding model that library was built with.

v4.9 - August 2, 2026

Documents no longer disappear after importing, and libraries carry their work between devices instead of asking you to redo it.

Your Libraries

  • Fixed documents disappearing shortly after import. A document that finished importing while the app was saving your library could be dropped from the list, even though it had imported correctly. This affected a single device as well as several, and it is what made the sample documents unreliable.
  • Documents processed on one device no longer need re-importing on another.
  • Libraries no longer appear to lose documents while iCloud is still catching up.

iPhone and iPad are coming from v4.7, so everything under v4.8 below ships here too. On Mac, v4.8 was already released, so only this section is new.

v4.8 - July 31, 2026

Deep Think and Maximum were not wired up correctly in previous releases. This release fixes that, and everything it uncovered.

Deep Think And Maximum

  • Both modes now reason across your documents. Previously they returned Standard quality answers after a much longer wait.
  • Answers cite their sources in every mode. Maximum produced none at all.
  • Both stop once they stop finding new material, typically halving Maximum’s run time.
  • A single failed pass no longer ends a query.
  • Resolved “The selected model isn’t available right now.”

Privacy And Routing

  • On-Device now covers the entire query, including the final answer.
  • The model picker governs every mode. It previously reached Standard only, so Deep Think and Maximum fell back to a default.

What You See

  • The live pipeline names each stage correctly, including verification and query rewriting.
  • Reasoning detail wraps instead of cutting off mid-sentence.
  • Follow-up suggestions come from the answer rather than stray words.
  • Raw model output no longer appears in answers.
  • Passes skipped for having no relevant text are shown instead of leaving gaps.
  • An answer reporting that your documents do not cover something is kept, not replaced with generic help text.

Also

  • Document import works on Mac. The file picker there was a placeholder.
  • Opening the app after an update now shows what changed.

v4.7 - July 28, 2026

  • Honest Route Labels: Labels now say exactly where each answer ran: on your device, or Apple Private Cloud Compute with your permission. Nothing claims more than the system can verify.
  • Steadier Document Understanding: The key ideas pulled from your documents now come out the same every time, for more consistent search and more reliable connections across files.
  • Route Verification: New internal checks confirm that every answer’s recorded route matches what actually ran.
  • Removed Unsupported Model Claims: Settings no longer advertises selectable 3B or 20B on-device models, or a parameter-count based capability chip. (Correction to earlier releases: the public SDK exposes no such selector. Apple’s larger on-device model is real and managed by the operating system, but no app can choose or observe it.)

v4.6 - July 15, 2026

  • Smarter Private Cloud Compute Decisions: Queries now retrieve evidence locally before choosing a model route. PCC is used only when the final evidence and context budget justify secure cloud synthesis; missing evidence never causes escalation.
  • Exact Cloud Consent: The confirmation sheet describes why PCC was selected and shows the size of the minimized evidence payload that would be transmitted. Background actions never wait on a hidden consent prompt.
  • Model Choice That Stays Put: Hybrid, On-Device, and PCC now describe the policy you selected, not whichever route happened to run last. Hybrid remains selected after a PCC answer and after relaunch.
  • Route Badges on Answers: Every Apple-model answer identifies the route that completed it: on-device, PCC, or on-device fallback. The badge opens the existing detailed route receipt.
  • Reliable Local Fallback: Hybrid and explicit PCC requests fall back on-device when consent, network, entitlement, quota, or PCC availability blocks cloud execution, provided streaming has not meaningfully begun. The answer is labeled as a fallback instead of silently changing routes.
  • Remembered PCC Choice: Always Allow and Never Allow persist across relaunches. The app no longer opens a generic PCC sheet on startup; permission is requested only for a real finalized evidence package.
  • Clear GPU Execution Profiles: Efficiency, Balanced, Performance, and Maximum replace the misleading percentage control and align Settings with the work the app can actually route to GPU-capable paths.
  • Accurate Route History: Saved response metadata distinguishes intended, attempted, actual, fallback, and completed execution paths.
  • Truthful Model Reporting: The “Advanced” on-device model preference now reports the actually executed model route in telemetry and diagnostics. (Correction to earlier releases: the current OS SDK exposes no separate 20B on-device model API; the Advanced preference executes the standard on-device model.)
  • Apple-Approved PCC Capability Enabled: The v4.6 source entitlement is enabled and the app verifies the signed process entitlement before constructing Apple’s native PCC model. Signed iOS 27 physical-device and TestFlight validation remains pending.
  • Hardened Knowledge-Index Migrations: Database schema migrations are driven by a fixed, code-owned migration catalog, eliminating a class of malformed-schema risk.
  • Ingestion Stop Now Persists: Closing or discarding an ingestion queue prevents those exact jobs from returning after iCloud reload. Automatic empty-index repair runs one library at a time and remains disabled for that library on the device where you dismissed it until you explicitly import or rebuild again.
  • Restored Test Coverage: The unit-test target removed in an earlier release is restored, with regression suites pinning embedding, parsing, citation, and launch-argument behavior.
  • PR Backlog Consolidation: All 43 open automated pull requests were audited end-to-end; the valuable changes were reimplemented cleanly in this release and the remainder documented and closed.

v4.5.1 - July 2, 2026

  • Parallel PDF Ingestion Concurrency Fix: Resolved concurrency race conditions and deadlocks on Apple Silicon by introducing thread-safe NSRecursiveLock serialization around CoreImage image generation, preventing concurrent Metal context crashes during multi-page parallel processing.
  • Silicon-Native Core AI Embeddings: Compiled and bundled the EmbeddingModel.aimodel format inside the package resources. This activates Apple’s zero-copy memory paths for 40%+ faster sentence embedding execution natively on Apple Silicon. Included a companion model compilation utility (scripts/compile_core_ai_model.py) to easily convert PyTorch model graphs.
  • Flexible Settings Saving: Enabled saving of embedding configuration options when a provider is unavailable at save-time, allowing runtime fallback routing (e.g. Core AI falling back to Core ML) to resolve and execute cleanly.
  • Silicon HUD Layout Correction: Restored the iOS Silicon HUD position back to its original layout coordinates (x: 45, y: safeAreaInsets.top + 85), and configured the HUD legend to dynamically shift to the right (x: 345) when the conversation history sidebar is visible to prevent overlap.
  • Private Cloud Compute Fallback & UI Safeguard: Configured dynamic Hybrid (Automatic) model routing to check for active developer entitlements before selecting Private Cloud Compute, ensuring seamless fallbacks to local models and eliminating routing delays. Greyed out and disabled the PCC option in the header model preference selection menu until the official developer entitlement is active.
  • Historical Advanced Picker Correction: The earlier RAM-gated “20B Advanced” label was not backed by a distinct public SDK model selector. v4.6 removes that active claim and migrates the saved preference to the public On-Device target.

v4.5 - July 2026

  • Rust-Backed Tokenizer Engine: Replaced the legacy pure-Swift tokenizer with a highly optimized Rust-backed swift-tokenizers engine. This delivers much faster document tokenization and exact byte-level character offset tracking for citations. (A “100x speedup” was quoted here originally; it was never measured and was withdrawn 2026-08-06.)
  • Ingestion Pipeline Stability: Fixed FTS5 index corruption and page offset mapping bugs during streamed ingestion of large documents. Large files are now fully searchable.
  • Core AI Diagnostic Options: Stabilized Silicon-native Core AI sentence embeddings on iOS 27+ / macOS 27+ with robust model readiness indicators and detailed build/OS error diagnostics in settings.
  • CI/CD Pipeline Upgrades: Updated cloud build environments to macos-26 to natively support Xcode 27+ and Swift 6.2+.

v4.4 - June 2026

  • Evidence Threads: Introduced durable conversational threads. Chat sessions and their active citations, metadata, and responses are now persisted on disk, eliminating dynamic ephemerality.
  • Slide-Out Thread Sidebar: Added a premium, spring-animated slide-out menu to view, create, switch, and delete research threads. Features full swipe-to-delete support and active-selection highlighting.
  • Design System Parity: Structured the new sidebar using the app’s native visual tokens (DSColors, DSSpacing, DSTypography), rendering a visually integrated, sleek interface on both iOS and macOS.
  • Thread Storage: Threads initially shipped in isolated local-only storage. (Updated July 2026: thread storage has since moved into the app’s workspace storage and now syncs across your devices via iCloud Drive using safe, coordinated file writes. If you edit the same thread on two devices at nearly the same time, the most recent change wins.)
  • Engineering Diagnostics View: Retained the debug view and mocked thread triggers inside the Developer Diagnostics Hub to inspect and verify thread persistence.
  • Native Private Cloud Compute Support: Integrates route-policy support and diagnostics to escalate complex queries to Apple’s secure Private Cloud Compute server enclaves on iOS 27 / macOS 27+ for supported and entitled builds, with graceful automatic fallback to local on-device execution when unavailable.
  • Pro Annual Pricing & Free Trial: Calibrated the Pro Annual subscription to $29.99/year (representing a 58% savings vs monthly) and introduced a 7-day free trial.
  • Discontinued Document Pack Add-on: Discontinued the consumable Document Pack add-on, removing related UI cards, quick-refill views, and purchase flows.
  • Direct Review Prompts: Streamlined the app-rating experience by triggering Apple’s native review prompt directly at key “happy moments” (like tapping Thumbs Up), eliminating the intermediate alert dialog.

v4.3 - June 20, 2026

  • Transparent Verification Engine: When the AI abstains from answering due to lack of evidence, it will now gracefully provide its drafted reasoning with a prominent [Needs Verification] warning, rather than hiding the drafted answer entirely.
  • Accurate Telemetry Status: Fixed a minor UI bug that caused the Verification Gates status panel to incorrectly light up red (as “Failed”) during perfectly healthy standard queries.
  • Instant Library Scrolling: Fixed a scroll-lag issue when browsing large libraries, by caching row identifiers instead of recomputing them on every row draw.
  • Faster Answer Generation: Rebuilt the context aggregation math so deduplicating large evidence sets no longer slows down as the library grows. Answers generate substantially faster.
  • Enhanced Reliability: Simplified the underlying AI routing engine to be exclusively reliant on native Apple Intelligence, resulting in fewer context errors.

v4.2 - June 2026

  • Modernized UI: Completely rebuilt the live telemetry dashboard with premium frosted glass and interactive haptic feedback.
  • Dynamic Verification Gates: The visual HUD for RAG telemetry now adapts its pipeline dynamically based on your active RAGQualityMode.
  • Fixed Chat History Persistence: Resolved an issue that sometimes skipped loading your previous chat history during a cold boot.
  • Granular Hardware Telemetry: The Execution Badge now dynamically fetches and displays exact onboard RAM allocations alongside TOPS processing power.
  • Accuracy in Retrieval Metrics: Corrected UI labels to differentiate between semantic Database Matching (Vector Similarity) and active LLM reasoning thresholds.

v4.0 & v4.1 - WWDC26 Apple Intelligence Update

OpenIntelligence version 4.0 & v4.1 is a major Apple Intelligence modernization and refinement pass currently live on the App Store. The release touches every major component of the user experience, from the local-first execution model to transparent citation details, visual evidence cards, and Siri/Shortcuts system integration.

This document consolidates this major release cycle into a single, cohesive user-facing log, highlighting the Apple Intelligence native foundation alongside a GPU-accelerated ingestion pipeline, Metal vector search performance, live reasoning telemetry, and database protection.

The practical user story is simple: OpenIntelligence now does a better job showing what evidence it used, where it ran, and how much source support it found before you trust an answer.


Why WWDC26 Matters Here

WWDC26 shifted the platform architecture by pushing core AI capabilities into system-level frameworks. For OpenIntelligence, this unlocked several important system resources:

  • Apple Foundation Models: Native LanguageModelSession instances, structured generation, token budgeting, and route-policy layers.
  • Apple Private Cloud Compute (PCC): A secure cloud route for complex or context-heavy work, ensuring cryptographic privacy.
  • App Intents & App Entities: Integration with Siri, Shortcuts, and system services for reading, listing, and indexing documents.
  • Visual Intelligence: A platform route to import OCR and camera captures as active RAG evidence.
  • Core Spotlight: Deep system-level indexing of document chunks and sections.
  • Core AI: Execution pathways for local custom models.
  • Liquid Glass: A modern visual design system featuring responsive glass effects.

What Users Will Notice First

  • Transparent Verification UI: Responses are explicitly labeled so you can see if they are Source-Locked (fully supported by your documents), Partially Supported, or Lacking Sufficient Evidence.
  • Model Routing Visibility: The status pill in the header shows you exactly where your query is targeted to run: On-Device (standard questions) or Private Cloud Compute (complex or long-context questions, falling back locally if the build lacks the required cloud entitlement).
  • Live Reasoning Telemetry: You can watch the model’s active thinking loop in real-time inside the bottom metrics bar as it organizes thoughts, resolves routing, and writes answers.
  • Lag-Free Visual Transitions: The processing dashboard and overlays transition smoothly using GPU-accelerated effects, eliminating visual stutter during document imports.
  • GPU-Accelerated Local Vector Search: Local vector similarity calculations run on Apple Silicon hardware accelerators using custom Metal pipelines. (A “4x” figure was quoted here originally; it was never measured and was withdrawn 2026-08-06.)
  • Clean Discarding & Deletion: Canceling or deleting an in-progress import now triggers a cascading purge that completely removes file fragments, database records, and Spotlight search indexes, ensuring no orphaned data is left behind.
  • Grammatically Correct Suggested Questions: Suggested follow-up questions are grammatically clean and diverse across different sections of your library.

Key Improvements

1. Smarter On-Device vs. Private Cloud Compute Routing

  • Dynamic Policy: Local execution is preferred for standard queries to protect battery life and latency. Heavy reasoning queries or large files are dynamically targeted to route to secure Private Cloud Compute, with automatic local fallback if the entitlement is not present.
  • Under the Hood Details: Tap the header status pill to open a popover detailing the active model name, token budget usage, and explanations of Apple’s PCC privacy guarantees.
  • Direct Route Metrics: The status bar displays telemetry from the actual resolved routing engine rather than an estimation based on response latency.

2. Grounded Answers and Citation Integrity

  • GroundedAnswerView: Presents cited answers clearly, making it easy to map each statement back to its specific source document.
  • Visual Evidence Cards: Image imports, camera scans, and PDF figures render inside the chat bubbles as OCR-derived evidence.
  • Fidelity Status: Clearly displays the verification level of every response, helping you decide how much to rely on the generated answer.

3. Better Recovery When Generation Misbehaves

  • Empty Response Fallback: If a model returns an empty response, the RAG service now routes into a reliability fallback instead of treating the entire query as unavailable.
  • Partial-Draft Preservation: If streaming produced a useful partial answer before a failure, the app preserves that text instead of replacing it with an empty or generic failure result.
  • Rate Limit Safeguards: Rate-limited or concurrent Apple Foundation Model failures get a short retry path before falling through to recovery behavior.
  • Stricter Grounding Repair: Stricter repair pathways handle context overflow and missing citations, prompting abstention when grounding support is too weak.

4. UI Telemetry & Render Optimizations

  • Thinking Stream: The ThinkingStreamView shows live feedback during reasoning phases so you are never left wondering if the app is frozen.
  • GPU-Driven Overlay: Ingestion overlay animations utilize hardware-accelerated opacity and scale transitions, preventing frames from dropping during heavy background indexing.
  • Massive Document Stability: Resolved a critical bug where opening the app after ingesting massive documents (e.g., HOA docs) caused an instant crash or UI freeze. Ingestion logs are now lazily loaded and capped to prevent unbounded memory allocation and excessive view generation on launch.

5. Suggested Questions & Grammar Safeguards

  • Diverse Suggestions: The query planner isolates unique sections of your documents to guarantee follow-up questions cover a variety of topics.
  • Grammar Filter: Uses Apple’s native language taggers (NLTagger) to analyze parts of speech and filter out OCR layout noise, verbs, or incomplete phrases from suggested prompts.

6. GPU & Neural Engine (ANE) Pipeline Optimizations

  • Adaptive Pre-Scan: Automatically pre-scans documents before ingestion. Clean digital PDF pages bypass expensive Vision OCR pipelines completely, achieving up to a 20% processing speedup.
  • GPU Resolution Scaling: Scales document rendering resolution dynamically based on font size risk, reducing document parsing memory requirements on device silicon.
  • Hardware Telemetry: Integrates haptic and visual metrics feedback for Vision OCR, vector embedding generation, and LLM inference.

7. Database Safety & Zero-Remnant Discarding

  • Anti-Corruption Writes: Vector database disk saves write to a contiguous memory buffer and replace database files atomically, preventing corruption if the app reloads during ingestion.
  • Zero-Remnant Purge: Deleting or discarding a document cleanly deletes all corresponding vector data, FTS5 database indices, Spotlight entries, and local files.
  • Sync Tombstones: Leverages deletion logs (deleted_documents.json) to prevent deleted documents from being revived during cross-device syncing.

8. System Integrations: Spotlight, Siri, and Shortcuts

  • Deep Spotlight Indexing: System Spotlight search can search and index down to specific document chunks, sections, and figures.
  • App Intents: Persisted document and library entities are exposed to Siri and Shortcuts. You can ask Siri to summarize, compare, or search documents directly.
  • Visual Intelligence OCR: Captures image inputs and extracts their text as active evidence in the RAG pipeline.

9. App-Wide UI, Onboarding, and Workflow Improvements

  • Onboarding Progress: Updated the checklist and imports dashboard to display live stages, extraction progress, vector generation counts, and a timer publisher for smooth elapsed-time tracking.
  • Live Activities: Integrated Live Activity support to show background import status directly on the lock screen and Dynamic Island.
  • Sample Document: Renamed the sample document to “OpenIntelligence Product Guide” to align with onboarding instructions.

10. Retrieval, Summaries, and Evaluation Quality Gates

  • Evaluations Framework: Built a suite to measure retrieval recall, citation precision, exact-value accuracy, and hallucination rates against strict quality gates before updates are shipped.
  • RAPTOR-Lite Routing: Added summary routing to handle high-level document overviews by querying generated document summaries.

11. Liquid Glass and Visual Polish

  • Universal AppIcon: Added a unified universal AppIcon configuration across iOS and macOS targets, resolving catalog build warnings.
  • Visual density: Standardized margins (14pt) and tighter corner radii for message bubbles (16pt) and cards (12pt) to create a denser, more cohesive Liquid Glass UI.
  • Model Preference Selector: The original selector is superseded by v4.6’s persistent Hybrid, On-Device, and PCC policies. PCC fallback is now explicitly labeled rather than represented as PCC running locally.