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 • On my Mac, "How much notice do I have to give before I move out?" came back "1 lb." from an air fryer manual, marked Verified. Any "how much" 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 that question 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. • When you ask how much something holds, a capacity written without a space, like "4.5L", counts as liters now. PLANS • 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, down from its $59.99 regular price, and it's just called Lifetime now. • 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, all from your own App Store's prices. Everything still runs on your device, unless you okay Apple's Private Cloud Compute.

This one's all about answers. They're done when the last word shows up, most Deep Think and Maximum answers write out in front of you now, and on iPhone the app tells you when one finishes while you're off doing something else. FASTER ANSWERS • Standard was sitting on answers that were already done. If you asked for one specific fact, it ran a second check after the answer was 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. Send your next question or hit Stop and the answer stays, sources and all. ANSWERS AS THEY'RE WRITTEN • Deep Think and Maximum 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 Standard answers stream too. They used to get written in one piece and then played back. WHEN YOU LOOK AWAY • On iPhone you get a light tap when an answer's done, plus a badge on the Chat tab if you were somewhere else. • The timer keeps running when you come back to an answer that's still going. It used to freeze at the moment you left. • If iOS cuts off the app's background time before an answer finishes, whatever it already wrote stays in the chat with a note saying why. Before, it just disappeared. Everything except that one writing step still runs on your device.

This release is about what you see first: the sample library's questions, the welcome screens, and the upgrade screen. Underneath it, documents are read with Apple's own detector and there is a new, off-by-default control over how answers are written. THE SAMPLE LIBRARY MAKES SENSE AGAIN • The three sample documents come with questions written by hand to walk you through the app. One leftover duplicate of a sample was enough to switch them off and replace them with questions built from templates, which is where "What is nothing?" came from. Duplicates are cleaned up on your next visit to Documents, and the written questions show whenever the three samples are there. • When Apple Intelligence has nothing well-grounded to ask about a document, the app now shows no suggestion rather than a bad one. READING YOUR DOCUMENTS • Addresses, phone numbers, dates, amounts of money, measurements, flight numbers and tracking numbers inside tables are now recognised by Apple's own detector instead of pattern matching that only understood US phone numbers. Prose on the same page does not yet contribute; that is the current limit, stated plainly. ANSWERING • Three sliders in Model Parameters said they changed how the model picks words. Apple's generation API has no penalty setting of any kind, so they never did, on the device or on Private Cloud Compute. They now say so. • New, and off until you turn it on: Adapt to the question. The app picks a temperature and a length from what you asked, so looking up a value gives the same number twice and comparing two things gets room to work. It applies in Standard, Deep Think and Maximum. It is off because nobody has shown it produces better answers yet; it is a setting you can judge for yourself. • You can write your own reasoning profile for the questions that reason on Private Cloud Compute, in your own words, in Model Parameters. THE WELCOME SCREENS AND THE UPGRADE SCREEN • The welcome screens follow your light or dark setting. In light mode the status bar was nearly invisible against them; it is not now. • A thumbs-up on an answer now offers to rate the app, write a review, or send feedback, once per version. The Lifetime card says how many months of Pro Annual its price buys, from the store's own prices. • While Lifetime is on sale, one line at the top of Chat and Documents says the real discount and the last day, computed from the store's own price. Dismiss it and it stays away. And once, after the fourth answer the app has verified against your files, it shows you the plans; it will not ask again. • The upgrade screen leads with what a plan actually changes. Every plan runs the same model, on your device, with Deep Think and consent-gated Private Cloud Compute included. Pro and Lifetime lift the daily cap on Maximum mode and raise the document and library limits. Pro Annual no longer offers a free trial; it is one payment a year. Everything except that one writing step still runs on your device.

This release turns on Private Cloud Compute. It needs iOS, iPadOS or macOS 27; on earlier systems everything continues to run on the device exactly as it did in 5.1. PRIVATE CLOUD COMPUTE • When a question needs more room than the model on your device can hold, the app can now hand that single writing step to Apple's Private Cloud Compute. Reading your files, searching them, choosing what to cite and checking the finished answer still happen on your device. • Nothing leaves without asking. Before anything is sent, a sheet shows you 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. • Measured on an iPhone with the A18 Pro: roughly 86 tokens per second from Private Cloud Compute, against 27 on the device. DEEP THINK AND MAXIMUM ASK FOR MORE • iOS 27 lets an app tell the server model how hard to think, and only one of the paths that write an answer was asking. The most common one was not, so those two modes often ran at ordinary effort while describing themselves as doing more. All of them ask now. • Expect both modes to take longer than they did in 5.1, and to use more of your daily Private Cloud Compute allowance. A CORRECTION FROM 5.1 • How It Works and the About screen told every iOS 26 user that the app asks before sending a question to Private Cloud Compute. Those builds could not send anything anywhere. Both screens now describe the build you are actually running, and so do the built-in guides, the Glossary and the Settings capability list. Everything except that one writing step still runs on your device.

Two releases went out on the Mac that iPhone and iPad never received. This one brings them across, along with the changes made since. OPENING AND MOVING AROUND • The Documents tab made you wait while it counted your cached documents. That count fed a single row which stays hidden unless the number is above zero, and on the device this was traced on it was always zero, so the row was never drawn. You waited for a number that was then thrown away. It now loads in the background and the tab opens straight away. • Switching libraries was the slow thing people actually noticed, and nothing on that path was being measured, so four separate attempts to find the cause had missed it. It is measured now, and faster. • "Analyzing corpus…" was an empty state wearing a progress spinner. It could never finish, because nothing was running behind it. THINGS THAT WERE TELLING YOU THE WRONG THING • The Semantic Atlas labelled a neuroscience paper "API Reference" and "Glossary". The fault was in three separate places, not one. • Turning the execution profile up removed hardware instead of adding it, and the control itself was a dropdown showing one option while hiding three, on a setting whose only purpose is choosing between them. • The Deep Think card described a minimum number of reasoning steps that does not exist. • Temperature stayed adjustable under a sampling mode that ignores it entirely. • The hardware panel reported a limit of 1,073,741,824 threads, which is 1024 cubed and not a real ceiling. It also now reports free memory rather than a number that looked like it but was not. DOCUMENTS • Refreshing one of the built-in samples left the old copy behind, so the library grew a duplicate every time you did it. • A tag that appears exactly once in a document no longer gets treated as a description of the whole thing. TEXT RECOGNITION • The app was asking the system to guess what language each page was written in, on every page, while also handing it a list of thirteen languages. Those two instructions contradict each other, and a wrong guess means your text is corrected against the wrong dictionary, which damages it rather than just slowing things down. The language is now worked out once, from the document itself. • Page images went to text recognition at the highest possible resolution with no downscaling, which is the slowest setting available. Recognition now scales to what is needed to read the smallest print a real document contains. IF YOU QUIT MID-IMPORT • A large PDF interrupted by quitting the app came back showing zero progress, as though the work was gone. It never was: the app resumes from the last page it finished. But the screen said otherwise, and removing the item is the one action that does discard that progress. A resumed import now tells you how far it got. Everything still runs on your device.

Documents were quietly losing parts of themselves, answers were built from a fraction of what was found, and the app rewrote your library on every launch. This is the fix for all three. SEARCH • The part of the app that reads what a passage means was looking at one position instead of the whole thing. It reads all of it now. • A broken length check returned the same number for every input, so long passages were cut short before they were ever indexed. ANSWERS • Deep Think was writing from a fraction of what it found, discarding its own best passages before the first word. It keeps them now, and says what it left out. • Citations are checked against the real source list. An answer could cite a source number past the end of its own list. • Deep Think is about three times faster, and stops when it runs out of new material instead of re-reading. YOUR DOCUMENTS • Tables in Word documents were read and then thrown away. A file could import looking complete with all of its numbers missing. • A tag that appears once no longer describes a document. Running headers were being turned into tags. YOUR LIBRARIES • A document you import no longer loses its searchability to the app's own housekeeping, which deleted a just-built index and said nothing. It now says so and rebuilds. • "Remove Local Copies" is now "Remove All Documents", because that is what it did. It deleted from iCloud and your other devices while saying Sync Now could bring them back. • Changing the embedding model no longer wipes your vectors before you agree to rebuild them. SPEED • The app starts faster. A 43 MB model loaded on every launch before anything appeared. • iCloud stopped re-uploading libraries that had not changed. One launch could rewrite hundreds of megabytes identical to what was there. • Leaving the chat no longer cancels the answer you were waiting for, and coming back keeps your place. • The Documents tab stopped waiting on a number it never showed you. It cost up to 393 milliseconds on every open. YOUR HARDWARE • Macs were being given iPhone-sized limits. Every Mac below the very top landed in a tier tuned for phones, and base M4 and M5 were demoted again on top of that, costing a 32 GB M5 six times its vector batch. • Turning performance up was making the app slower at the thing it does most. The four profiles were inverted, so the fastest setting gave the slowest import. • The performance selector is four cards, each saying which parts of the chip it engages. It was a dropdown hiding three. • The app understands Apple chips that do not exist yet. An iPhone 17 or M5 no longer reports itself as "A12 or Older", and anything newer scales forward instead of falling back. • Rotating no longer leaves black rectangles around the floating hardware readout, which now also shows free memory. SETTINGS • Settings is a searchable list instead of one long scroll. • Choose Top-K, Top-P or Greedy. The app used to decide, and always picked the same one. • Temperature is disabled, with a reason, on the one sampling mode that ignores it. • Five switches that never controlled anything are no longer switches. THINGS WE WERE CLAIMING THAT WEREN'T TRUE • Settings listed eight agentic tools and all eight were wrong. Four are wired up. Those four are what it names now. • We removed the claim that answers can run on Apple's Private Cloud Compute today. Shipped builds do not contain it yet. • Pages, Numbers and Keynote were advertised and never worked. They are no longer advertised, and importing one now fails clearly. • The hardware panel claimed a ceiling of 1,073,741,824 threads. That is 1024 cubed. It was multiplying three limits instead of reading one. • Deep Think said "6 to 8 reasoning sessions". There is no minimum; it stops when it has what it needs. • The Atlas was labelling a medical paper "API Reference", matching fragments instead of whole words. • "Analyzing corpus..." was never analyzing anything. It was a spinner that could not finish.

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

This update makes OpenIntelligence more precise about what it tells you. - 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. - The key ideas pulled from your documents now come out the same every time, for steadier search and more reliable connections across files. - New internal checks make sure every answer's recorded route matches what actually ran. Same goal as always: ask hard questions of your own files, see exactly where every answer came from, and trust what you can verify, not what you're told.

v4.6 is a major Apple Intelligence routing, transparency, reliability, and jargon-reducing update. EVIDENCE-FIRST MODEL ROUTING: - OpenIntelligence now searches your library before deciding where an answer should run. The decision uses the evidence actually found, the amount of context required, and whether the question needs synthesis across multiple documents. - Normal work remains on-device. On supported iOS, iPadOS, and macOS 27 systems, longer evidence-backed requests can use Apple's native Private Cloud Compute when entitlement, availability, network, quota, and consent requirements are satisfied. - Missing or weak evidence never triggers cloud escalation. The app can stay local or abstain when the library does not contain enough information. CLEARER CONSENT AND EXECUTION HISTORY: - Private Cloud Compute consent now happens after the final evidence package is prepared. The confirmation view explains why PCC was selected and shows the number of source passages, context size, and estimated payload size. - Saved response details now distinguish the intended route, attempted route, route that actually ran, any fallback, and the route that completed the answer. - Automatic mode can fall back on-device before meaningful response text has streamed. Cloud and local partial answers are never stitched together, and Cloud Only requests report why they could not run instead of silently changing routes. - Model labels and diagnostics now reflect the Apple model route the public SDK actually executed instead of claiming a selectable model tier that the operating system does not expose. PRIVATE CLOUD COMPUTE SUPPORT: - Enabled OpenIntelligence's approved native PCC capability for supported builds and added platform-specific entitlement checks for iPhone, iPad, and Mac. - Availability and quota are checked again immediately before a PCC session begins. Unknown or unavailable quota states fail safely and do not authorize cloud execution. - Retrieval, evidence selection, citation checking, and final verification remain on-device even when PCC performs the final synthesis step. INGESTION THAT RESPECTS STOP: - Closing or discarding an ingestion queue now records that decision during iCloud reconciliation, preventing those exact jobs from reappearing after a workspace reload or stale sync snapshot. - Automatic repair of an empty document index now runs one library at a time and stays disabled for a dismissed library on that device until you explicitly import or rebuild again. - Cancellation waits for a safe document boundary so stopping a repair does not leave half-removed catalog metadata behind. INDEXING AND EXTRACTION RELIABILITY: - Knowledge-index migrations now use a fixed, code-owned migration catalog with stricter identifier validation, reducing the risk of malformed database upgrades. - Document chunk metadata has a deterministic fallback when system language tagging cannot return named entities, keeping retrieval useful in asset-constrained environments. - Expanded regression coverage now protects model routing, fallback behavior, embeddings, citation parsing, structured answers, ingestion recovery, and launch configuration. OpenIntelligence is still focused on the same goal since its inception: help users question complex source material, understand how each answer was produced, and return to the evidence when verification is imperative.

Version 4.5 introduces native Rust-backed text processing, memory-safe document streaming, optimized GPU calculations, and Apple Intelligence integrations. 1. Rust-Backed Tokenizer and Citation Precision: - Rust-Backed Tokenizer Engine: Moved text tokenization from Swift string-splitting loops to a pre-compiled native Rust library (swift-tokenizers) wrapped in a local package, accelerating document tokenization and index compilation by up to 100x. - Precision Character Offsets: Utilizes native Rust-calculated byte-level character offsets to provide high-precision source citations back to original text passages. - Isolated Resource Bundling: Moved tokenizer resource directories entirely to local package targets to prevent Xcode synchronized group resource duplicate copy warnings and flattening conflicts. 2. Memory-Safe Ingestion and Processing: - Bounded Stream Processing: Re-engineered document parsing from whole-file loading to a memory-safe streaming pipeline. This prevents RAM spikes and mitigates Out-of-Memory (OOM) risks on large PDF documents, keeping RAM usage below 32MB. - Zero-Copy Vision Extraction: Bypassed traditional PNG encoding overhead during document ingestion by routing GPU-rendered pages directly into the Vision OCR engine, accelerating text extraction by over 30%. - Page Offset Tracking: Corrected page-offset mapping during extraction to align database references with printed page numbers for accurate citation lookup. - Incremental Search Indexing: Upgraded the SQLite FTS5 index to append data on each streaming iteration, resolving index truncation bugs and retaining full-text search records across streaming batches. - Ingestion Queue Protection: Added a 15-minute file modification age check and mutated storage relative path mappings to prevent synchronization daemons from sweeping active files. - Live Ingestion Telemetry: Refined the Dynamic Island and Apple Watch Smart Stack widget layouts to display progress rings and processing percentages during background document imports. 3. GPU-Accelerated Similarity and Core AI: - Metal 4 Similarity Engine: Accelerated vector database matching by running SIMD4 and threadgroup-level calculations directly on the Metal GPU engine, speeding up query matching by up to 4x. - Dynamic Core AI Routing: Defaults sentence embedding tasks to Apple's native Core AI framework on iOS 27+ / macOS 27+ devices, while maintaining backward-compatible CoreML fallbacks for iOS/macOS 26. - Private Cloud Compute Fallbacks: Added intelligent entitlement verification for Private Cloud Compute queries on iOS 27. If secure cloud permissions are pending or unavailable, reasoning-heavy queries gracefully and transparently fall back to local on-device Foundation Models without crashing. - Single-Instance Caching: Caches the model provider instance to prevent double allocation during startup, saving up to 100MB of system RAM. - Model Pre-Warming: Integrated background pre-warm APIs to pre-load Apple Intelligence foundation models in 0.01 seconds, ensuring instant first-query responsiveness. 4. Billing System and Diagnostics: - StoreKit 2 Re-alignment: Re-engineered StoreKit product queries and entitlement loops. Removed deprecated document-pack consumables and stabilized entitlement reconciliation across subscription tiers (Monthly, Annual, and Lifetime Cohort). - Telemetry Diagnostics: Added a dedicated diagnostics panel inside the settings pane to show compile-time and runtime model readiness status. - Configuration Resilience: Refined the AI Subsystem settings panel to allow saving provider configurations even when the target hardware is temporarily warming up or unavailable, ensuring smooth runtime fallback routing.

Version 4.4 introduces persistent Evidence Threads with iCloud synchronization, refined Siri voice shortcuts, and enhanced self-healing background processing. 1. Durable Evidence Threads and iCloud Sync: - Persistent Conversations: Replaced ephemeral chat histories with durable research threads stored locally under the system Application Support directory. Citations, metrics, and generated responses are preserved across app restarts. - iCloud Drive Synchronization: Implemented coordinated bidirectional synchronization of research threads across user devices using iCloud Drive, protected by file-coordination safeguards to prevent write conflicts. - Monetization Tier Quotas: Thread creation is managed using tier-specific quotas (5 for Free, 20 for Pro, and Unlimited for Lifetime subscribers), with clear, localized notifications when quota thresholds are reached. - Isolate Active Threads: Updated the "New Chat" action to allocate a new active thread rather than deleting the previous conversation from disk, isolating active threads per container to prevent workspace bleed. 2. Advanced Siri and Shortcuts Integration: - Split Settings Interface: Rebuilt the settings panel to separate Siri voice integrations from Shortcuts automation libraries. - Siri Voice Shortcuts: Displays the 9 pre-registered voice command shortcuts mapped to active thread and search capabilities. - Shortcuts Actions Library: Displays the 16 custom App Actions available for drag-and-drop workflows in the system Shortcuts App, categorized by task (Document Ingestion, Retrieval, Summarization, History, and Diagnostics). - In-Process Intent Execution: Refactored the Shortcuts App Actions. Intents now resolve directly on the active running user interface instance via a weak static reference holder (activePresentedInstance), causing the presented views to reload and switch instantly. - Library Entity Parameters: Added optional library container parameter support to App Intents, letting automated workflows target specific document silos instead of always falling back to default folders. 3. RAG Engine Refinements and Model Constraints: - Fuzzy Phrase Verification: Refined the anti-hallucination verification gate logic to prevent false-positive refusals by performing fuzzy plural/singular word mappings and ignoring query-specific auxiliary terms. - Ungrounded Fallback Compliance: Updated the Standard and Reliability mode pipelines to respect ungrounded fallback preferences, preventing unnecessary response discarding when ungrounded output is explicitly allowed. - Hardened On-Device Constraints: Hardened model preference routing to strictly route execution on-device and cap RAG context packing budgets (6K tokens) when the local 3B Core or 20B Advanced models are selected, completely bypassing Private Cloud Compute boundaries. - Swift 6 and Diagnostic Polish: Resolved all Swift 6 compiler warnings and concurrency actor isolation errors. Adjusted the Quick Sanity Check to prevent false similarity failures across CoreML and legacy embedding providers. Added a direct external link to the Notion Roadmap database to the About Screen. 4. Compatibility Requirements: - OS Version Compatibility: Runs on macOS 26.x and iOS 26.x. Advanced zero-copy Silicon-native sentence embeddings under Core AI (.aimodel) and native Private Cloud Compute secure enclaves require macOS 27 or iOS 27. - Hardware Compatibility: Offline Neural Engine model features are optimized for Apple Silicon (M-series and A17 Pro+ or later). 5. IAP Changes - Calibrated Pro Annual to $29.99/year (saving 58% vs monthly) and introduced a 7-day free trial.

Changes since WWDC 2026: - Fixed Image Playground - Unleashed RAM ceilings for exponential pipeline scaling on advanced Apple Silicon, and precisely aligned context limits to the official AFM 3 API specs (4K/32K). - Powered by the AFM 3 Architecture: Updated model configuration parsing to dynamically route and visibly highlight across the entire third-generation model suite (AFM 3 Core, AFM 3 Core Advanced, and AFM 3 Cloud Pro). - Screen Awareness: Integrated AppIntents background context frameworks to allow Siri to natively ingest on-screen files and URLs directly into RAG libraries without touching the app. - ADM 3 Cloud Integration: Plumbed Apple's ADM 3 architecture via Image Playground API into the core generation pipeline for instant visual concept rendering. - Lightning-Fast Answer Generation: Rebuilt the RAG deduplication pipeline using O(N) Set-based tracking, resulting in an over 1,000x speedup in evidence aggregation for large libraries. - Buttery-Smooth Database Dashboard: Implemented a dynamic UUID dictionary cache in DatabaseDashboardView, accelerating row rendering performance by ~240x during heavy scrolling. - Removed legacy OnDeviceAnalysisService to simplify LLM routing, fully trusting Apple Intelligence native FoundationModels. - Integrated the 20B Apple Foundation Model (AFM 3 Core Advanced) into the execution pipeline. Prioritized .onDeviceAdvanced routing over Private Cloud Compute to maximize local privacy and eliminate cloud latency for reasoning operations. - Hardened token budget obedience and evaluation suites to maintain extreme robustness against Apple Intelligence constraints. - Streamlined configuration parsing by removing legacy strictMode boolean from KnowledgeContainer, mapping directly to native minSimilarity thresholds. - Pruned visual noise by removing obsolete logic relating to the deprecated "Fibonacci sphere" distribution in AdaptiveVisualizationsView. - Expanded OpenIntelligenceEngineTests suite with hybrid search safety checks and Semantic Chunker hardening against empty strings and malformed data. - Fixed agentic reasoning orchestration so that Standard mode strictly prohibits auto-escalating to Deep Think loops when utilizing the constrained 3B Core model. - Reclassified VerificationGateService Domain Isolation gate to an advisory status to prevent abstention false-positives on cross-domain queries. - Hardened BackgroundTaskService against BGTaskSchedulerErrorDomain error 3 by eliminating string dynamic identifiers and wildcards from system registration logic. - Harmonized all availability macro targeting across the entire codebase to iOS 26.0, macOS 26.0, correctly aligning with Apple's 2025 unified naming architecture. - Fixed string interpolation in AppIntents parameter summaries. - Dropped 3 experimental iOS 27.0 AppShortcuts to strictly enforce Apple's 10-shortcut system limit. --- Version 4.2: - Modernized UI for macOS/iOS: Completely rebuilt the live telemetry HUD utilizing iOS 26/macOS 26/WWDC26+ APIs. Integrated .ultraThinMaterial for premium glassmorphism, hardware .sensoryFeedback for interactive haptics, and smooth .symbolEffect animations. - 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 after force-closing the app. --- Version 4.1 & 4.0: - Apple Foundation Models Integration: Migrated language model sessions to native iOS 26+ FoundationModels APIs. Deconstructed the monolithic LLM services into dedicated helper modules. - Dynamic Routing Policy: Standard queries route to local on-device models (4K token context boundary), while complex or long-context queries automatically scale to secure Private Cloud Compute (32K tokens). - Core AI Local Scaffolding: Staged CoreAISentenceEmbeddingProvider as experimental local scaffolding under Apple's Core AI framework.

Changes since WWDC 2026: - Unleashed RAM ceilings for exponential pipeline scaling on advanced Apple Silicon, and precisely aligned context limits to the official AFM 3 API specs (4K/32K). - Powered by the AFM 3 Architecture: Updated model configuration parsing to dynamically route and visibly highlight across the entire third-generation model suite (AFM 3 Core, AFM 3 Core Advanced, and AFM 3 Cloud Pro). - Screen Awareness: Integrated AppIntents background context frameworks to allow Siri to natively ingest on-screen files and URLs directly into RAG libraries without touching the app. - ADM 3 Cloud Integration: Plumbed Apple's ADM 3 architecture via Image Playground API into the core generation pipeline for instant visual concept rendering. - Lightning-Fast Answer Generation: Rebuilt the RAG deduplication pipeline using O(N) Set-based tracking, resulting in an over 1,000x speedup in evidence aggregation for large libraries. - Buttery-Smooth Database Dashboard: Implemented a dynamic UUID dictionary cache in DatabaseDashboardView, accelerating row rendering performance by ~240x during heavy scrolling. - Removed legacy OnDeviceAnalysisService to simplify LLM routing, fully trusting Apple Intelligence native FoundationModels. - Integrated the 20B Apple Foundation Model (AFM 3 Core Advanced) into the execution pipeline. Prioritized .onDeviceAdvanced routing over Private Cloud Compute to maximize local privacy and eliminate cloud latency for reasoning operations. - Hardened token budget obedience and evaluation suites to maintain extreme robustness against Apple Intelligence constraints. - Streamlined configuration parsing by removing legacy strictMode boolean from KnowledgeContainer, mapping directly to native minSimilarity thresholds. - Pruned visual noise by removing obsolete logic relating to the deprecated "Fibonacci sphere" distribution in AdaptiveVisualizationsView. - Expanded OpenIntelligenceEngineTests suite with hybrid search safety checks and Semantic Chunker hardening against empty strings and malformed data. - Fixed agentic reasoning orchestration so that Standard mode strictly prohibits auto-escalating to Deep Think loops when utilizing the constrained 3B Core model. - Reclassified VerificationGateService Domain Isolation gate to an advisory status to prevent abstention false-positives on cross-domain queries. - Hardened BackgroundTaskService against BGTaskSchedulerErrorDomain error 3 by eliminating string dynamic identifiers and wildcards from system registration logic. - Harmonized all availability macro targeting across the entire codebase to iOS 26.0, macOS 26.0, correctly aligning with Apple's 2025 unified naming architecture. - Fixed string interpolation in AppIntents parameter summaries. - Dropped 3 experimental iOS 27.0 AppShortcuts to strictly enforce Apple's 10-shortcut system limit. ----------- Version 4.2: - Modernized UI for macOS/iOS: Completely rebuilt the live telemetry HUD utilizing iOS 26/macOS 26/WWDC26+ APIs. Integrated .ultraThinMaterial for premium glassmorphism, hardware .sensoryFeedback for interactive haptics, and smooth .symbolEffect animations. - 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 after force-closing the app. ----------- Version 4.1 & 4.0: - Apple Foundation Models Integration: Migrated language model sessions to native iOS 26+ FoundationModels APIs. Deconstructed the monolithic LLM services into dedicated helper modules. - Dynamic Routing Policy: Standard queries route to local on-device models (4K token context boundary), while complex or long-context queries automatically scale to secure Private Cloud Compute (32K tokens). - Core AI Local Scaffolding: Staged CoreAISentenceEmbeddingProvider as experimental local scaffolding under Apple's Core AI framework.

Version 4.2: - Modernized UI for macOS/iOS: Completely rebuilt the live telemetry HUD with ultra-thin materials, interactive haptics, and smooth symbol animations. - Dynamic Verification Gates: The visual HUD for RAG telemetry now adapts its pipeline dynamically based on your active Quality Mode. - 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 (Total Confidence). - Agentic Tool Visibility: The telemetry HUD now surfaces implicit, hidden engine calls (such as Vector Search Engine lookups) during standard modes that do not trigger recursive tool event streams. - Clean Sub-second Telemetry: Fixed a visual bug presenting Time-To-First-Token in raw oversized milliseconds (e.g., 12000ms), standardizing values to under 1.2s formatted duration. - Native Resizable Telemetry Drawer: Completely rebuilt the expanded metrics panel to behave like a fluid, native iOS bottom sheet. Users can now physically pull the handle down to manually resize the metrics view seamlessly during live telemetry inspection or recording. ----------- Version 4.1: - Core AI Sentence Embedding Provider: Added CoreAISentenceEmbeddingProvider for high-speed, local silicon-accelerated vector calculations on Apple device hardware. - Real-Time Thinking Telemetry: Added ThinkingStreamView inside the UnifiedMetricsBar to show live reasoning and model processing states. - Enhanced Suggested Questions: Upgraded SuggestedQuestions with a two-pass section-diversity selector and strict POS-tagging grammar filters to generate high-quality follow-ups. - Metal GPU-accelerated retrieval: Integrated a custom Metal compute shader pipeline with SIMD4 and threadgroup-level execution, driving a 4x speedup in batch cosine similarity RAG retrieval. - Atomic Database & Purging: Hardened RAG pipeline with atomic vector database writes and cascading file/index deletion of discarded uploads to prevent data corruption. - Thread-Safe Safety Routing: Thread-safe MainActor routing for LLM availability checks. ----------- Version 4.0: - Dynamic Model Routing (On-Device & PCC): Automatically routes queries based on complexity and context size. Standard queries execute locally using the 4K-token on-device model, while complex logic scales to Private Cloud Compute (PCC) enclaves supporting a 32K-token context window. - Under the Hood Telemetry Dashboard: Added an interactive details popover detailing active model routing, token budgets, execution path telemetry, and pulsing status indicators. - Core AI Engine Integration: Integrates a direct, custom local Core AI silicon execution engine and model registry. - Native Liquid Glass UI: Re-engineered key view components with premium native glasscard modifiers for an immersive visual experience. - RAG Evaluations Suite: Built-in dataset validation against target Recall@5 and Citation Precision quality gates, exposing an Apple Evaluations Bridge for native CLI testing compatibility. - Agentic Retry Safeguard: Hardened agentic RAG reasoning cycles to preserve non-empty drafts and protect against rate-limit empty responses. - Brand Realignment: Realigned all branding assets, HUD panels, and logs to standardized "Apple Intelligence" styling.

Changes since 3.7.5: Version 4.1: - Core AI Sentence Embedding Provider: Added CoreAISentenceEmbeddingProvider for high-speed, local silicon-accelerated vector calculations on Apple device hardware. - Real-Time Thinking Telemetry: Added ThinkingStreamView inside the UnifiedMetricsBar to show live reasoning and model processing states. - Enhanced Suggested Questions: Upgraded SuggestedQuestions with a two-pass section-diversity selector and strict POS-tagging grammar filters to generate high-quality follow-ups. - Metal GPU-accelerated retrieval: Integrated a custom Metal compute shader pipeline with SIMD4 and threadgroup-level execution, driving a 4x speedup in batch cosine similarity RAG retrieval. - Atomic Database & Purging: Hardened RAG pipeline with atomic vector database writes and cascading file/index deletion of discarded uploads to prevent data corruption. - Thread-Safe Safety Routing: Thread-safe MainActor routing for LLM availability checks. Version 4.0: - Dynamic Model Routing (On-Device & PCC): Automatically routes queries based on complexity and context size. Standard queries execute locally using the 4K-token on-device model, while complex logic scales to Private Cloud Compute (PCC) enclaves supporting a 32K-token context window. - Under the Hood Telemetry Dashboard: Added an interactive details popover detailing active model routing, token budgets, execution path telemetry, and pulsing status indicators. - Core AI Engine Integration: Integrates a direct, custom local Core AI silicon execution engine and model registry. - Native Liquid Glass UI: Re-engineered key view components with premium native glasscard modifiers for an immersive visual experience. - RAG Evaluations Suite: Built-in dataset validation against target Recall@5 and Citation Precision quality gates, exposing an Apple Evaluations Bridge for native CLI testing compatibility. - Agentic Retry Safeguard: Hardened agentic RAG reasoning cycles to preserve non-empty drafts and protect against rate-limit empty responses. - Brand Realignment: Realigned all branding assets, HUD panels, and logs to standardized "Apple Intelligence" styling.

Changes since 3.7.5: Version 4.1: - Core AI Sentence Embedding Provider: Added CoreAISentenceEmbeddingProvider for high-speed, local silicon-accelerated vector calculations on Apple device hardware. - Real-Time Thinking Telemetry: Added ThinkingStreamView inside the UnifiedMetricsBar to show live reasoning and model processing states. - Enhanced Suggested Questions: Upgraded SuggestedQuestions with a two-pass section-diversity selector and strict POS-tagging grammar filters to generate high-quality follow-ups. - Folder Picker Persistence: Implemented security-scoped file bookmark directory mapping, resolving iCloud Drive and File Provider path sandboxing constraints. - Atomic Database & Purging: Hardened RAG pipeline with atomic vector database writes and cascading file/index deletion of discarded uploads to prevent data corruption. - Thread-Safe Safety Routing: Thread-safe MainActor routing for LLM availability checks. Version 4.0: - Dynamic Model Routing (On-Device & PCC): Automatically routes queries based on complexity and context size. Standard queries execute locally using the 4K-token on-device model, while complex logic scales to Private Cloud Compute (PCC) enclaves supporting a 32K-token context window. - Under the Hood Telemetry Dashboard: Added an interactive details popover detailing active model routing, token budgets, execution path telemetry, and pulsing status indicators. - Core AI Engine Integration: Integrates a direct, custom local Core AI silicon execution engine and model registry. - Native Liquid Glass UI: Re-engineered key view components with premium native glasscard modifiers for an immersive visual experience. - RAG Evaluations Suite: Built-in dataset validation against target Recall@5 and Citation Precision quality gates, exposing an Apple Evaluations Bridge for native CLI testing compatibility. - Agentic Retry Safeguard: Hardened agentic RAG reasoning cycles to preserve non-empty drafts and protect against rate-limit empty responses. - Brand Realignment: Realigned all branding assets, HUD panels, and logs to standardized "Apple Intelligence" styling.

Changes since 3.7.5: - Dynamic Model Routing (On-Device & PCC): Automatically routes queries based on complexity and context size. Standard queries execute locally using the 4K-token on-device model, while complex queries (so long as it's selected in the Settings tab) scales to the now accessible Private Cloud Compute (PCC) enclaves supporting a 32K-token context window. - Under the Hood Telemetry Dashboard: Added an interactive details popover detailing active model routing, token budgets, execution path telemetry, and pulsing status indicators. - Core AI Engine Integration: Integrates a direct, custom local Core AI silicon execution engine and model registry. - Native Liquid Glass UI: Re-engineered key view components with premium native glasscard modifiers for an immersive visual experience. - RAG Evaluations Suite: Built-in dataset validation against target Recall@5 and Citation Precision quality gates, exposing an Apple Evaluations Bridge for native CLI testing compatibility. - Agentic Retry Safeguard: Hardened agentic RAG reasoning cycles to preserve non-empty drafts and protect against rate-limit empty responses. - Brand Realignment: Realigned all branding assets, HUD panels, and logs to standardized "Apple Intelligence" styling. If you have any questions, suggestions, or want to know more - please send me an email using the Feedback feature! Also, please leave a review!!

Changes since 3.6: Version 3.7.5: - Hardened iCloud synchronization concurrency by offloading all synchronous database merging, file copying, and conflict resolutions off the Main Actor (UI Thread) to completely eliminate UI freezes and watchdog crashes. - Accelerated multi-device data transfers using parallel Swift TaskGroups to download iCloud ubiquitous files concurrently rather than sequentially. - Hardened download error recovery, ensuring isolated iCloud file download failures or timeouts do not stall or fail the entire library synchronization. Version 3.7/3.7.1: - Resolved a gesture conflict on iOS where long-pressing library pills in the horizontal scroll view failed to trigger the context menu, fully restoring library deletion on iPhones. - Preserved full library identity in the Documents pill strip more reliably by preventing the document-count badge from collapsing names into ambiguous truncation. - Fixed a synchronization issue in iCloud Sync where deleted libraries could be merged back and resurrected on other devices, and implemented deletion tombstones to automatically propagate deletions across all synced devices. - Added automated local cleanup of vector databases, Spotlight search indexes, and UI presentation caches when a synced library is deleted on another device. - Documents was tightened again with cleaner library pills, a less crowded header, smaller sync controls, and clearer organization/management surfaces. - Shared-workspace and background-ingestion plumbing are more robust, with safer queue cleanup, cleaner reconciliation, and better handling for long-running work. - Camera capture, OCR-heavy pages, and mixed digital/scanned documents import more reliably. - Clean digital text is preserved more faithfully, while noisy scans and image-heavy pages still get the heavier recovery path when they need it. - Retrieval is stronger across Standard, Deep Think, and Maximum, with better context packing, better use of surrounding document context, and less drift away from the source. - Suggested questions and follow-ups are more grounded in the active library and less repetitive across refreshes. - Chat handles direct attachments and captured content more smoothly, so it is easier to bring new material into the conversation flow. - Answer inspection is much richer now, with clearer source review, timing, retrieval-quality, and evidence-detail surfaces when you want to see how a response was built. - Technical answers and structured output render more cleanly, including stronger code block handling and clearer response detail views. - Diagnostics and device-aware performance behavior are more stable on larger libraries and longer-running work, with deeper inspection tools behind the scenes for validation and monitoring. - Added native App Store rating and review prompting triggers after successful query tasks. - Resolved Mac Catalyst layout truncations, including the Sync Mode picker, action chips, and scrollable library selector pills. - Enabled full iCloud ubiquity container access and network permissions for Mac Catalyst by packaging universal sandbox entitlements. - Resolved Xcode build catalog warnings with a unified universal AppIcon configuration across iOS and macOS targets. - Redesigned the Silicon hardware telemetry HUD to dynamically rotate motherboard borders (SoC and Taptic outlines) to match device layout rotation, added iPad layout coordinates, and cleanly hid visual outlines on Mac targets. - Hardened suggested questions and 3D visualization keywords to aggressively filter out OCR junk, syntax noise, and generic templates. This release is about making OpenIntelligence feel more complete from import to answer review: fewer weak spots between "I added a file" and "I trust this answer."

Changes since 3.6: Version 3.7/3.7.1 is a broader release that tightens almost every stage of the app: library management, import, retrieval, answer quality, chat ergonomics, and diagnostics. - Resolved a gesture conflict on iOS where long-pressing library pills in the horizontal scroll view failed to trigger the context menu, fully restoring library deletion on iPhones. - Preserved full library identity in the Documents pill strip more reliably by preventing the document-count badge from collapsing names into ambiguous truncation. - Fixed a synchronization issue in iCloud Sync where deleted libraries could be merged back and resurrected on other devices, and implemented deletion tombstones to automatically propagate deletions across all synced devices. - Added automated local cleanup of vector databases, Spotlight search indexes, and UI presentation caches when a synced library is deleted on another device. - Documents was tightened again with cleaner library pills, a less crowded header, smaller sync controls, and clearer organization/management surfaces. - Shared-workspace and background-ingestion plumbing are more robust, with safer queue cleanup, cleaner reconciliation, and better handling for long-running work. - Camera capture, OCR-heavy pages, and mixed digital/scanned documents import more reliably. - Clean digital text is preserved more faithfully, while noisy scans and image-heavy pages still get the heavier recovery path when they need it. - Retrieval is stronger across Standard, Deep Think, and Maximum, with better context packing, better use of surrounding document context, and less drift away from the source. - Suggested questions and follow-ups are more grounded in the active library and less repetitive across refreshes. - Chat handles direct attachments and captured content more smoothly, so it is easier to bring new material into the conversation flow. - Answer inspection is much richer now, with clearer source review, timing, retrieval-quality, and evidence-detail surfaces when you want to see how a response was built. - Technical answers and structured output render more cleanly, including stronger code block handling and clearer response detail views. - Diagnostics and device-aware performance behavior are more stable on larger libraries and longer-running work, with deeper inspection tools behind the scenes for validation and monitoring. - Added native App Store rating and review prompting triggers after successful query tasks. - Resolved Mac Catalyst layout truncations, including the Sync Mode picker, action chips, and scrollable library selector pills. - Enabled full iCloud ubiquity container access and network permissions for Mac Catalyst by packaging universal sandbox entitlements. - Resolved Xcode build catalog warnings with a unified universal AppIcon configuration across iOS and macOS targets. - Redesigned the Silicon hardware telemetry HUD to dynamically rotate motherboard borders (SoC and Taptic outlines) to match device layout rotation, added iPad layout coordinates, and cleanly hid visual outlines on Mac targets. - Hardened suggested questions and 3D visualization keywords to aggressively filter out OCR junk, syntax noise, and generic templates. This release is about making OpenIntelligence feel more complete from import to answer review: fewer weak spots between "I added a file" and "I trust this answer."

Changes since 3.6: Version 3.7 is a broader follow-up release that tightens almost every stage of the app: library management, import, retrieval, answer quality, chat ergonomics, and diagnostics. If 3.6 made it easier to move your libraries across devices, 3.7 is the update aimed at making everything that happens after that feel sharper: importing harder files, attaching new material in chat, getting better-grounded answers, and having much clearer proof when the app answers confidently. - Documents was tightened again with cleaner library pills, a less crowded header, smaller sync controls, and clearer organization/management surfaces. - Shared-workspace and background-ingestion plumbing are more robust, with safer queue cleanup, cleaner reconciliation, and better handling for long-running work. - Camera capture, OCR-heavy pages, and mixed digital/scanned documents import more reliably. - Clean digital text is preserved more faithfully, while noisy scans and image-heavy pages still get the heavier recovery path when they need it. - Retrieval is stronger across Standard, Deep Think, and Maximum, with better context packing, better use of surrounding document context, and less drift away from the source. - Suggested questions and follow-ups are more grounded in the active library and less repetitive across refreshes. - Chat handles direct attachments and captured content more smoothly, so it is easier to bring new material into the conversation flow. - Answer inspection is much richer now, with clearer source review, timing, retrieval-quality, and evidence-detail surfaces when you want to see how a response was built. - Technical answers and structured output render more cleanly, including stronger code block handling and clearer response detail views. - Diagnostics and device-aware performance behavior are more stable on larger libraries and longer-running work, with deeper inspection tools behind the scenes for validation and monitoring. - Added native App Store rating and review prompting triggers after successful query tasks. - Resolved Mac Catalyst layout truncations, including the Sync Mode picker, action chips, and scrollable library selector pills. - Enabled full iCloud ubiquity container access and network permissions for Mac Catalyst by packaging universal sandbox entitlements. - Resolved Xcode build catalog warnings with a unified universal AppIcon configuration across iOS and macOS targets. - Redesigned the Silicon hardware telemetry HUD to dynamically rotate motherboard borders (SoC and Taptic outlines) to match device layout rotation, added iPad layout coordinates, and cleanly hid visual outlines on Mac targets. - Hardened suggested questions and 3D visualization keywords to aggressively filter out OCR junk, syntax noise, and generic templates. This release is about making OpenIntelligence feel more complete from import to answer review: fewer weak spots between "I added a file" and "I trust this answer."

Changes since 3.5: Version 3.6 adds optional iCloud reuse for the libraries you choose, without giving up the app's local-first default. If you've been loading a large library on iPad and wishing that exact processed library could show up on iPhone or your other devices without starting over, this is the update aimed at that problem. Shoutout to Tim for asking for this. - Every library can now be set to Local Only or iCloud Drive individually. - Local Only libraries stay fully on-device unless you explicitly change them. - Libraries you mark iCloud Drive can reuse imported files and processed state across your own Apple devices on the same iCloud account. - Shared-library refresh and review is clearer now, so additions or removals from another device are easier to understand before you pull them in or remove them here too. - Same-name shared libraries are much less likely to merge together unexpectedly. - iCloud library sync is now positioned as a paid workspace feature, and paid plan capacity is clearer: Pro supports up to 10 libraries and Lifetime supports up to 20. - If a long-running import is interrupted on one device, another device can pick up queued work for that iCloud library instead of forcing a full restart. - Plain text and other digital documents are preserved more faithfully during import, with safer handling for text, markdown, code, CSV, transcripts, and Office-style files. - This follow-up 3.6 build also smooths the Documents layout, improves import cancellation, and prevents old queued work from deleted libraries from coming back unexpectedly. This release is about making cross-device reuse practical without compromising the app's privacy-first, local-by-default model.

Changes since 3.2.5: Sorry for the rough edges in the last few updates. Version 3.5 is the cleanup release that should have landed sooner. If dense PDFs, exact-value lookups, starter prompts, or long-running imports felt less reliable than they should have, this is the corrective pass. It rolls up the real fixes shipped after 3.2.5 and makes the app more dependable on hard documents. - Exact answers are stronger across Standard, Deep Think, and Maximum for source-backed questions over tables, specs, measurements, counts, dates, prices, and similar exact values. - Starter questions and follow-ups are more grounded and less likely to surface weak, generic, or misleading prompts. - Onboarding, empty states, and the bundled sample workspace explain the app more clearly, including best-supported file types, the 4,096-token model limit, and when processing stays on-device versus uses Apple Private Cloud Compute. - PDFs and images now share one adaptive visual-ingestion path, searchable figures and structured tables survive more often, and clean scientific PDFs are less likely to produce fake tables, broken headings, or reference-section noise. - Table-heavy pages are less likely to collapse back into scrambled paragraph text during ingestion, improving retrieval quality after re-import. - Large user-initiated imports are more reliable, with better queue recovery, background cleanup, and long-running import handling. - Library and settings surfaces now describe per-library isolation and live runtime behavior more accurately. This release is about making OpenIntelligence more grounded, more predictable, and more trustworthy on real documents.

Version 3.2.5 What's changed since 3.2: - Exact fact lookups are much more direct. If the answer is clearly present in a table, spec row, or short source passage, OpenIntelligence now locks onto that source value faster instead of overthinking it. - Deep Think and Maximum now run a precision lookup before longer reasoning, so simple questions can still get quick cited answers even in the higher-effort modes. - Standard, Deep Think, and Maximum share stronger table/spec retrieval rescue for measurements, capacities, limits, prices, counts, and other exact values. - Suggested starter questions are now generated from actual passages instead of loose document labels, so "Ask something real" should be much more relevant to the library you uploaded. - The answer text for exact measurements is cleaner, including nearby equivalent units when the source provides them. - This is a fast corrective update for answer-quality regressions in the 3.2 line. Sorry for the quick follow-up, and thank you for bearing with the pace while I tighten the engine.

Version 3.2 What's changed since 3.1: - Deep Think and Maximum are more stable during repeated use, including back-to-back deep reasoning in the same thread and “Go Deeper” after a Standard answer. - A live reasoning strip now appears in the unified bar above chat so users can watch the current retrieval or reasoning step in real time if they want the extra transparency. - Long Deep Think answers are less likely to cut off mid-bullet or mid-thought, with better continuation handling when a response stops at a real boundary problem instead of a natural finish. - Standard got a higher-quality pass for non-trivial prompts, with adaptive multi-session reasoning for harder questions while still staying quick on obvious simple lookups. - This is another fast follow-up because 3.1 improved a lot of the pipeline but still left some answer-quality regressions in the live app. This build is the corrective pass focused on stability, completeness, and trust.

Version 3.1 Reliability upgrade focused on hard technical documents, noisy OCR, multi-column layouts, and grounded answer quality. - Deep Think and Maximum now preserve strong grounded partial answers when a late-stage generation interruption happens, instead of dropping generic stop text onto useful output. - Multi-column PDFs and corrupted tables now ingest more cleanly with layout-aware OCR fallback and better row and column preservation. - Table retrieval is stronger for specs, measurements, and statistical values through better schema, row, and cell anchoring. - Weak first-pass retrieval now triggers a corrective evidence pass before answer generation when the initial evidence pack is too thin or too generic. - Unsupported or weakly supported claims are handled more conservatively before final answers are shown. - Dense scientific PDFs and technical manuals hold up better during source review, with clearer structured excerpts and better abstention when support is weak.