OpenIntelligence

Apple Intelligence-fueled agentic RAG for iPhone and iPad. Import private libraries, run grounded retrieval through Apple's Foundation Models framework and the local 3B-parameter on-device model path, and inspect cited answers instead of generic chat output.

Shipping Apple Intelligence App Plus Engine Lineage

OpenIntelligence is the public app surface for a broader Apple Intelligence document stack, with an agentic retrieval engine boundary living behind the product.

OpenIntelligence ingests PDFs, Office files, code, images, audio, and video into private libraries, then runs an agentic retrieval-augmented generation loop over that material with query planning, retrieval routing, citations, confidence checks, and deliberate abstention when the evidence is weak. Most ingestion, indexing, retrieval, and Apple Intelligence answer orchestration stay local; when Apple-managed cloud capacity is used, it is only through platform-controlled paths rather than generic third-party AI hosting.

What Makes OpenIntelligence Different

Apple Intelligence on device

Apple's Foundation Models framework sits at the center of the answer path, with the local 3B-parameter on-device model route handling grounded work by default unless the user explicitly moves into an Apple-managed cloud path.

Broad document ingestion

PDFKit, Vision OCR, structured parsers, and media extraction paths handle digital documents, scanned pages, code, images, audio, and video.

Agentic RAG orchestration

SQLite FTS5, local vector indexing, query analysis, reranking, routing, and context packing are orchestrated as a real agentic retrieval loop instead of a single-vector-only search path.

Cited answers and abstention

The app is built to show evidence, drop unsupported claims, and refuse when the documents do not support a confident answer.

Library and container isolation

Private libraries, retrieval controls, and diagnostics make it possible to scope work and inspect how the answer path behaved.

App and engine split

The public product ships as an app, but the underlying document intelligence boundary is also treated as a reusable engine inside the broader OpenIntelligence line.

Technical Profile

Core APIs

  • UI: SwiftUI on iPhone and iPad
  • Parsing: PDFKit, Vision OCR, structured document processing
  • Models: Apple Intelligence via Apple's Foundation Models framework and the local 3B-parameter on-device path
  • Retrieval: agentic RAG with SQLite FTS5, local vector indexing, reranking, and evidence packing

Data boundary

  • Local: Libraries, normalized text, retrieval state, citations, and Apple Intelligence answer orchestration
  • Optional: Apple-managed cloud capacity only through supported platform paths
  • Not the product: generic chatbot behavior over arbitrary cloud context
  • Status: shipping public app plus broader engine lineage

How OpenIntelligence Runs

1

Import into private libraries

Files are assigned to local libraries and normalized through text extraction or OCR, depending on document quality.

2

Build lexical and vector indexes

The app preserves structure where possible, then stores lexical signals and vector representations for hybrid retrieval.

3

Plan, retrieve, and verify evidence

Query planning, reranking, routing, and evidence packing are used to keep the answer path tied to source material rather than generic model guesswork.

4

Use Apple Intelligence to answer or abstain

Answers are rendered with citations and confidence cues through the Apple Intelligence path, or the app abstains when support is not strong enough.

Support and Privacy Routes