Find exact answers
Search specifications, dates, measurements, table values, and requirements buried across large files.
Open Source · Built on Apple Intelligence
Ask your files anything. Inspect every answer.
OpenIntelligence turns your PDFs, Office and iWork documents, spreadsheets, presentations, notes, scans, code, and audio or video recordings into searchable libraries on iPhone, iPad, and Mac — then answers with citations you can tap to verify. Retrieval and verification always run on your device, and on-device answering works in airplane mode with no account and no API keys. Long, evidence-heavy requests can use Apple's Private Cloud Compute — only after the app shows you exactly what would be sent and you approve it — and every answer carries a badge showing the route that actually ran. Under the hood: an open-source 29-step RAG engine built on Apple's foundation models, with Vision OCR ingestion and Metal-accelerated retrieval.
No sign-up · No API keys · On-device answering works offline · Open source
App Store 4.6 · 4.8★ from 4 ratings · App Store privacy label: Data Not Collected
What can you do with it?
Understand complex material
Summarize reports, research papers, manuals, contracts, class notes, and meeting recordings.
Compare multiple sources
Ask questions across an entire library and inspect which files and passages support the result.
Search more than PDFs
Import documents, spreadsheets, presentations, code, images, scans, audio, and video — with Vision OCR and on-device speech transcription where needed.
Verify the answer
Tap citations to return directly to the supporting source instead of trusting an unexplained AI response.
Want to see how that works technically? Keep scrolling.
App Capabilities
Execution Policy Control
Choose Hybrid, On-Device, or Private Cloud Compute as a persistent policy. Routing decisions happen after retrieval, based on the evidence actually found — and every answer shows a receipt-backed badge for the route that completed it.
Universal Apple Silicon
Seamlessly synced and optimized across iOS, iPadOS, and macOS 26 and newer. A true native Swift engine built for iPhone, iPad, and Mac.
Citations & Verification
Verification Gates check each answer against the retrieved evidence, and every answer includes interactive citations that link directly to the source document.
Multi-Modal Ingestion
A native multi-modal pipeline that automatically extracts and generates embeddings from complex PDFs, DOCX files, and raw visual data via Vision framework.
RAPTOR Agentic Loops
Stitches together multiple reasoning sessions utilizing the RAPTOR Summary Router to overcome token limits and synthesize exhaustive answers.
Native Local RAG
Powered by SQLite FTS5 and Accelerated by threadgroup-level Metal pipelines across the CPU, GPU, and Neural Engine (ANE).
Real-World Execution
See It In Action.
Watch OpenIntelligence ingest files, retrieve evidence, generate an answer, and verify its citations directly on Apple hardware.
On-Device Technical Trace
How It Works Under the Hood.
The full engine is a 29-step pipeline: a 6-step ingestion lane plus a 23-step agentic query loop. Here is a live simulation of the engine processing local documents directly against the foundation model — the trace expands several steps into sub-stages so you can inspect each operation.
Apple Foundation Model Agentic Pipeline
OpenIntelligence Quality Modes
Pipeline Strategy
Execution Target
-- ms
Engine Latency
--
Throughput / Rate
--%
Platt Confidence
Simulation Speed
800ms
Step --
Select a node to inspect
WHAT IT DOES
Click on any node in the architectural diagram above to inspect how that processing step operates inside the native Swift engine.
WHY IT MATTERS
Understanding each phase of on-device RAG is critical to optimizing memory, latency, and context limitations.
HOW IT WORKS
Select a node to see technical implementation details, Swift API methods, and algorithmic settings.
Runtime Log Trace
No active log trace.
Engine Implementation Source
// Select a step to inspect source code.Public Roadmap
What's Next
This roadmap is synced automatically from our Notion database, giving you full transparency into what is currently being built and what is completed.
To Do
In Progress
Completed
Product History
Version Changelog
Read the official version release log for OpenIntelligence, focusing on user-facing capabilities and system architectural shifts.
4.6
v4.6 — Evidence-First Routing & Native PCC
July 2026The current App Store release. The app searches your library before deciding where an answer runs, and Apple approved the native Private Cloud Compute entitlement.
- Evidence-First PCC Routing: Post-retrieval planning decides the route from the evidence actually found; weak evidence never triggers cloud escalation.
- Per-Answer Route Badges: Hybrid / On-Device / PCC are persistent policies, and every answer shows the receipt-backed route that completed it.
- Post-Payload Consent: PCC permission is requested only after the exact cloud payload is final, with its size and reason shown.
- Honest Model Labels: Labels reflect the route Apple's public SDK actually executed — no selectable model-tier claims.
4.5
v4.5 & v4.5.1 — Core AI Embeddings & Streaming Ingestion
July 2026Silicon-native embeddings, a 100x faster tokenizer, and memory-safe ingestion for very large documents.
- Core AI Embeddings: Sentence embeddings on Apple's Core AI framework (iOS/macOS 27+) with Core ML fallback.
- Rust-Backed Tokenizer: ~100x faster tokenization with exact byte-level offsets for precise citation mapping.
- Streaming Large-PDF Ingestion: Page-level checkpointing that survives restarts, plus fixes for parallel-ingestion races and Apple Silicon deadlocks.
- AI Subsystem Diagnostics: X-ray view of the active model, Neural Engine target, tokenizer, and latency profile.
4.4
v4.4 Release — Evidence Threads
June 2026Approved today by Apple App Store review. Introduces persistent iCloud-synchronized research chat sessions, slide-out sidebar, and resolved Swift 6 compiler concurrency checks.
- iCloud Evidence Threads: Persistent conversation sessions saved locally and synced automatically via iCloud Drive.
- Sidebar Navigation: Slide-out menu panel for switching, managing, and deleting active chat sessions.
- Swift 6 Concurrency: Resolved actor isolation warnings inside key orchestrators.
- Entitlements Alignment: 1,000 document upload hard limit for Pro tiers and adjusted Annual subscription to $29.99/year.
4.3
v4.3 & v4.3.1 — AFM 3 Suite & Deadlock Fixes
June 2026Delivered support for third-generation Apple Foundation Models, Siri Screen Awareness AppIntents, and resolved critical MainActor deadlocks during iCloud file locking.
- AFM 3 Architecture Routing: Real-time local vs. cloud routing between 3B Core, 20B Advanced, and Cloud Pro models. (Superseded in v4.6: Apple's public SDK exposes no selectable model tiers, so routing is now expressed as Hybrid / On-Device / PCC policies.)
- Siri Screen Awareness: Siri background ingestion allows screen files/URLs to load natively into RAG libraries.
- Image Playground Integration: Bound native macOS/iOS Image Playground (ADM 3) APIs for visual summary rendering.
- Thread Concurrency Fixes: Offloaded synchronous operations and File Coordinator locks to background tasks.
4.2
v4.2 Release — Liquid Glass Telemetry HUD
June 2026Rebuilt the RAG telemetry HUD with glassmorphic layouts, hardware sensory haptics, and resizable metrics bottom sheets.
- Liquid Glass HUD: Telemetry overlay styled with
.ultraThinMaterialglassmorphism. - Dynamic Verification Gates: Adaptive UI paths tracking 4, 8, or 12 verification gates depending on query modes.
- Resizable Bottom Sheet: Fluid iOS sheet that can be manually expanded or collapsed.
4.0
v4.0 & v4.1 — Apple Intelligence Milestone
June 2026Major milestone release integrating native iOS 26+ FoundationModels APIs, Core AI embeddings, and Metal-accelerated vector search.
- FoundationModels Integration: Migrated LLM orchestration to Apple's native system APIs.
- Private Cloud Compute Routing (entitlement-gated): Built the routing layer targeting PCC enclaves; at the time, execution awaited Apple entitlement approval, so all queries ran on device. (Apple approved the entitlement in July 2026; native PCC shipped in v4.6.)
- Core AI Native Embeddings: Integrated
CoreAISentenceEmbeddingProviderfor zero-copy memory layouts on Version 27+ devices. - Metal GPU Vector Search: Threadgroup-level Metal acceleration for 4x faster vector search.
Straight Answers
Frequently Asked Questions
Does my data leave my device?
Not unless you approve it. Document ingestion, embedding, search, and answer verification always run on your device using Apple's foundation models, and answers are generated on-device by default. On iOS, iPadOS, and macOS 27 and newer, a long evidence-heavy request can use Apple's Private Cloud Compute for the final synthesis — but only after the app shows you exactly what would be sent and you give permission, and every answer displays a badge showing the route that actually ran. On older OS versions, every query runs entirely on device.
Does OpenIntelligence work offline?
Yes. On-device retrieval and answering run entirely on your device, so you can ingest documents and ask questions in airplane mode. Only the optional Private Cloud Compute route requires a network connection, and the app falls back on-device when there is none.
Do I need an account or API key?
No. There is no sign-up and no API key to configure. OpenIntelligence uses only Apple Intelligence — no third-party AI services.
What file types can it read?
PDFs, DOCX files, and images or scans via Apple's Vision OCR, processed through a native multi-modal ingestion pipeline.
Is OpenIntelligence open source?
Yes. The full Swift source, including the 29-step RAG engine (6 ingestion steps plus a 23-step query loop), is public on GitHub, so the privacy claims can be independently verified.
Built by Gunnar Hostetler · Support and privacy at Gunzino · More projects at Gunnarguy
Ready to ask your files anything?
Free on the App Store · iPhone, iPad, and Mac · No account, no API keys