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.

Version 3.0 Major reliability upgrade focused on hard technical documents, noisy OCR, and broken table extraction. Since 2.6: - Deep Think and Maximum now preserve strong grounded partial answers when a late-stage generation hiccup interrupts the final polish pass, instead of dropping a generic failure footer onto an otherwise useful response. - Repaired row alignment for corrupted and multi-column tables so extracted evidence is less likely to collapse into jumbled garbage. - Added a corrective retrieval pass that automatically re-scans chunk and page indexes before generation when first-pass evidence is weak. - Improved extractive handling for statistical tables, specification sheets, and dense factual lookups with stronger table-priority evidence selection. - Tightened grounded abstention behavior when retrieval is off-topic or structurally weak, reducing confident answers built on bad evidence. - Improved source review for dense scientific PDFs with better evidence packing and more reliable structured excerpts. - Added audit visibility for corrective retrieval so retrieval failures are easier to diagnose and verify.

Messed up version control - current is now 2.6! All Changes Since v2.1: - Updated App Description to better reflect what the app actually does. - Starter questions now come from representative samples of your active library, so the opening prompts better match the documents and sections you're actually working with. - Refreshing starter questions does a better job surfacing different grounded prompts instead of repeating the same generic ideas. - Follow-up suggestions are more reliable after each answer and when switching libraries, including deeper follow-up, clarification, and comparison flows. - Answers now go through stronger grounded-QA routing before they are shown, using a stricter path for fact-heavy questions and a more constrained synthesis path when needed. - Evidence verification is tighter before final answer presentation, so the app is less willing to bluff when support is weak. - Weakly supported answers are handled more clearly, with better answer review and source inspection when you need to check what was actually grounded. - Response details are easier to inspect, with better source cards, clearer filenames, and scrollable excerpts for dense passages. - Tables, lists, quotes, headings, separators, and other structured technical output render much better. - Malformed links in generated answers are repaired more aggressively. - PDF imports do a better job filtering garbage text and noisy extraction on messier files. - Importing of Audio, Video (transcription), Code, and Office has been unified across chat interface + libraries. - Settings, billing, and plan messaging were cleaned up to better match the actual entitlement model.

Version 2.1.1 NEURAL ANSWERS - Better source-backed answers with clearer source review - Better handling when the app does not have enough evidence to answer BETTER REVIEW - Clearer answer review and source inspection in chat - Improved behavior with larger or messier files PRODUCT COPY - Cleaner onboarding, settings, and App Store wording

Version 2.1.1 SETTINGS & ABOUT - Added a public product hub with direct links to the roadmap, feedback board, changelog, GitHub repo, and App Store listing - Cleaned up About and Settings messaging so in-app copy matches the current product more closely LIFETIME COHORT - Lifetime plan messaging now focuses on concrete limits and one-time purchase terms - Added a Lifetime banner in Settings with quick links to the product hub and changelog CHAT POLISH - Refreshed the quality mode picker so Standard, Deep Think, and Maximum feel more visually distinct - Added clearer mode labels like Fastest, Iterative, and Full sweep BILLING & APP STORE - Updated App Store metadata and in-app billing copy to better describe what OpenIntelligence actually does - Added export compliance declaration for smoother App Store uploads What’s next? - Looking into OpenClaw/Claude integration as well as many many many many others. Question for users: What would you like to see?

Version 2.1 WHAT YOU'LL NOTICE Smarter Answers - The AI now tells you when your documents don't cover a question instead of making something up - Topical mismatch detection: when your question barely overlaps with retrieved content, Evidence-First mode activates — the AI states what evidence exists instead of hallucinating - 3 new verification gates (7 total, A-G): Semantic Grounding checks the answer matches sources, Quote Faithfulness checks quoted text actually appears, Generation Quality catches repetitive output - Critical gates can trigger full abstention — the app refuses to answer rather than answer wrong Deep Think & Maximum Mode - Fixed app freeze when sending a second question while one is processing (two reasoning chains were fighting over Apple FM) - Cancel-and-replace: old query cancels cleanly, new one takes over - Cancellation checks at all 17 LLM call sites — no wasted processing Suggested Questions - Fixed prompt contamination — hardcoded Kia Sportage examples were biasing every library toward car questions - Questions now reference specific details from your actual documents - Switching libraries instantly clears stale suggestions — no more flash of wrong questions - Natural conversational tone instead of robotic templates Cross-Library Isolation - Answers no longer leak between libraries when switching mid-conversation (session ID + 8 guard checks) - Entity index, full-text search, and suggestions fully scoped per library - Deleting a library properly cleans all cross-references Onboarding Rebuilt - 2-page flow: welcome cards → live pipeline theater showing your docs being processed in real time - Metrics dashboard: words, chunks, vectors, processing time - 3 built-in sample docs (Pricing, RAG Architecture, Apple Intelligence) — bypass free tier limit - Haptic feedback on all 6 interactions, VoiceOver labels throughout, reduce motion support - Retry button when sample import fails AI Hub Results - Key Facts, Step-by-Step, Plain English, What's Missing, Illustrate now render with full markdown (bold, bullets, headers) - New Share button alongside Copy and Insert - Sheet starts half-height, draggable to full Other Fixes - Alert dialogs can be dismissed again (was stuck due to .constant() binding) - "Show All Insights" now shows a proper list (was an empty TODO) - Image Playground no longer biased toward car themes (removed Ford Mustang few-shot example) - Settings: dynamic confidence threshold, duplicate Vietnamese removed, gates label corrected to A-G - About screen shows actual version instead of hardcoded 1.0.0 UNDER THE HOOD Safety - 37 force-unwrap crash sites eliminated across 26 files (guard let, if let, default values) - 10 fatalError() calls → graceful URL.temporaryDirectory fallbacks - StoreKit stream continuation crash fixed (IUO → AsyncStream.makeStream) - 3 silent Vision catch blocks now log properly Dead Code - 1,278 lines removed: SettingsRootView (701), DeveloperSettingsView (350), OpenAIResponsesAPIService (178), VecturaVectorDatabase (53), TipKit overlays - 15 test files deleted (2,686 lines) — were never wired into a test target Search Internals - BM25Scorer: class → struct with Storage reference type, fresh NLTokenizer per call (thread safety), lemmatization removed, division-by-zero guard - Font cipher detection prevents 93% content loss on encoded PDFs (Jaccard similarity) - Raw regex fix (5 patterns), garbled image extraction fix, dynamic image text budget Concurrency - Swift 6 strict concurrency: 11 files with nonisolated(unsafe) annotations, zero runtime change - GPU mode relabeling: Performance→Maximum, Balanced→Performance, new Balanced tier Developer - 25+ views gained #Preview macros for Xcode canvas - Pipeline X-Ray Studio: new 5,540-line web app for visualizing pipeline traces (dev tool, not in app) Includes all v2.0 features: 29-step RAG pipeline, AI Hub, 3-tier Metal GPU, device-specific OCR, markdown rendering, 102 services.

Version 2.0 AI-POWERED DOCUMENT TRANSFORMS New AI Hub toolbar with 5 document-aware transforms -- each uses your actual retrieved source chunks, not just the AI response text. - Key Facts: Source-backed bullet points with document/page attribution - Step-by-Step: Real procedures with specs and part numbers from your docs - Plain English: Simplifies complex technical content into accessible language - What's Missing?: Identifies gaps between your question and the retrieved answer - Illustrate: Image Playground visualization powered by LLM concept extraction SEARCH & RETRIEVAL UPGRADES - BM25 length normalization aligned for better recall on uniform-size chunks - 5 regex patterns pre-compiled as static constants - NLTagger lemmatization in keyword search - Anti-hallucination Gate E uses hardware-optimized cosine similarity - Verification Gate thresholds tuned to reduce false abstentions IMAGE PLAYGROUND On-device LLM translates domain jargon into visual scenes. No more try another description on technical content. FORMATTED RESPONSES Full markdown rendering: headers, bullets, bold, code blocks, block quotes. 6 regex patterns normalize Apple FM single-line output. PERFORMANCE (per Apple Silicon chip) - 2-4x parallel cross-encoder reranking via TaskGroup - 3x faster MLMultiArray fills with dataPointer bulk copy - Device-specific OCR concurrency (A19 Pro 8 ops, A18 Pro 6, A17 Pro 4) - 50-80% OCR skip rate on clean digital PDFs - 3-tier Metal GPU: threadgroup/SIMD4/scalar auto-selected per query size - 5 OCR candidates evaluated per recognition (was 3) RELIABILITY - Compression capped at 5 chunks with fresh LLM session per chunk - 12-second compression time budget - Rate-limit retry with 2s backoff - Extractive fallback upgraded: 6 chunks x 500 chars with section titles ZERO-DATA-LOSS INGESTION - Font cipher detection prevents 93% content loss on font-encoded PDFs - 5 Unicode regex patterns fixed - Image batch reduced from 20 to 5 pages for lower memory usage STABILITY - MMR crash fixed (GPU matrix edge case) - StoreKit 5s timeout (was 30-60s hang offline) - BGTask registration fixed - Swift 6 concurrency: 11 files updated, zero runtime change - 102 services, 175+ tests across 15 files

Version 1.2.0 RICH MARKDOWN RESPONSE RENDERING Responses now display with full formatting — headers, bullets, numbered lists, bold, code blocks, and block quotes. • Block-Level Markdown Parser — h1-h6, bullet lists, numbered lists, code fences, block quotes, horizontal rules, paragraphs. • Inline Normalization — Apple FM concatenates markdown on one line. 6 regex patterns split headers, bullets, numbered items onto separate lines before parsing. • Response-Cleaning Functions Audited — Two functions were stripping ALL markdown. Both rewritten to preserve formatting while removing artifacts. • Formatting-Aware LLM Prompts — All system prompts now instruct the model to use headers, bullets, and bold. • Paragraph Structure Preservation — Sentence joining uses paragraph breaks, maintaining document structure. DEVICE-OPTIMIZED PERFORMANCE ENGINE Every pipeline stage adapts to your specific Apple Silicon chip. • 3-Tier Metal GPU Search — Automatically selects fastest shader: Threadgroup (≥1000 vectors), SIMD4 (100-999), or Scalar fallback. Previously all searches used the slowest kernel. • Device-Specific OCR Concurrency — A19 Pro: 8 ops @ 1ms. A18 Pro: 6 @ 2ms. A17 Pro: 4 @ 3ms. Older: 2 @ 6ms. • Adaptive OCR Page Filtering — Pre-screens pages, skipping 50-80% of clean digital pages before OCR runs. • Concurrent GPU Rendering — CIFilter preprocessing now runs in parallel (was serialized). • Concurrent Neural Reranking — Cross-encoder scores pairs via TaskGroup. Bulk dataPointer copy instead of per-element subscript — 3× faster. • GPU Embedding Ingestion Mode — Uses CPU+GPU compute units during ingestion, freeing Neural Engine for OCR. • 5-Candidate OCR — Evaluates 5 transcription alternatives per recognition (was 3). STABILITY & HARDENING • Fixed crash in MMR diversification (dimension 0 edge case) • Matrix dimension validation added to RAGEngine • GPUComputeService edge cases — empty arrays no longer wrapped in spurious outer arrays • StoreKit product loading races against 5-second timeout — no more 30-60s hangs on airplane mode • Fatal threading assertion replaced with safe fallback in Release builds • Pipeline trace diagnostics compiled out of Release via #if DEBUG • Privacy manifest updated with SystemBootTime API declaration • Dead code removed from cross-encoder path ZERO-DATA-LOSS INGESTION Prevents silent content loss on font-encoded PDFs. • Font Cipher Detection — PHASE -1 validation detects PDFs where characters are shifted (Kia, Hyundai manuals). Jaccard similarity between OCR and PDFKit text. Garbled docs route through full Vision OCR. Prevents 93% content loss. • Raw String Regex Fix — 5 Unicode patterns were failing in Swift raw strings. CJK bullet, en-dash, em-dash patterns now work. • Garbled Image Extraction Fix — Image extraction on font-encoded PDFs no longer skips diagrams. • Dynamic Image Text Budget — Scales with document complexity instead of fixed limit. SWIFT 6 COMPLIANCE • 11 files updated with strict concurrency annotations. Zero runtime behavior change.

Version 1.1 • Updated descriptions and metadata to accurately reflect the 23-step RAG pipeline. • Clarified on-device storage and privacy guarantees. • Documentation improvements.