OpenIntelligence: Ask Your Files and Follow Every Citation

OpenIntelligence starts with a practical question: can you ask a library of personal files something useful without uploading the library to a chatbot? The app imports documents, creates a private index on Apple hardware, and returns answers with citations that open the supporting passages. Its strongest promise is not “AI knows everything”; it is that you can inspect the evidence and see when your files do not contain an answer.

From files to evidence

The supported inputs include PDFs, Office documents, code and plain text, images, audio and video. Vision OCR helps when a PDF’s text layer is unreliable. Pages, Numbers and Keynote files need to be exported first. The app can also work through very large PDFs using streamed, batched import with page checkpoints, according to the project documentation.

For a question, OpenIntelligence combines meaning-based vector search with keyword search, merges and reranks the candidates, then assembles a diverse evidence set. Each citation points back to an excerpt and page so you can check whether a sentence really follows from the source. When retrieval is weak, the app says so rather than treating a fluent answer as proof.

There are three quality modes. Standard is the direct, one-pass option. Deep Think runs a series of reasoning sessions over rotating context windows, while Maximum raises the session ceiling for questions spanning a larger library. The project reports one measured Standard baseline: 0.410 exact match on 83 QASPER cases from 40 research papers. That is a specific benchmark result, not a universal accuracy percentage; Deep Think and Maximum do not yet have scores on that same set.

What “local-first” means here

The repository says document ingestion, OCR, embeddings, indexing, retrieval, ranking and answer verification run on the device. The App Store description likewise says the files are searched locally first and that no third-party AI provider sits in the core path. On iOS/macOS 26, the documented behavior is local-only.

For iOS/iPadOS/macOS 27, both the App Store’s version 5.3 notes and the repository describe an optional Apple Private Cloud Compute (PCC) route for the final writing step. The app is supposed to show the selected evidence excerpts and ask before sending; retrieval and citation checking remain local. It is not a switch to an arbitrary third-party AI service.

There is an important release-documentation conflict: the App Store’s 5.3 listing says PCC is enabled, but the current GitHub README says App Store builds use the newest released Xcode and compile the PCC code path out. So treat PCC availability in the downloaded binary as unconfirmed until you see the route and consent controls on your own supported device. The local search-and-citation workflow is the dependable baseline.

Install from the App Store

The current US listing is OpenIntelligence on the App Store. Apple’s lookup reports version 5.3, released on 19 September 2026; the download is free with in-app purchases. The listing requires iOS/iPadOS/macOS 26 or later and describes Apple Intelligence-capable hardware: iPhone 15 Pro or later, or an M1-or-later iPad or Mac. The Store lists English as the app language.

The GitHub Releases page currently tops out at v4.9.0, older than the App Store build. Use the App Store link above for an end-user installation; the repository is the place to inspect the source and engineering notes, not the preferred download route.

License and who it suits

OpenIntelligence’s source is MIT-licensed. That makes the code reusable under the license’s notice conditions, but it does not relicense Apple frameworks, the App Store service or bundled third-party components. The app itself is free to download and offers in-app purchases; check the Store for current plan terms rather than relying on old release notes.

It is a good fit for researchers, students and professionals who repeatedly look up facts across manuals, papers, contracts or code and want to verify the source. It is less compelling if you need a fully localized interface—the Store currently lists English—or expect every quality mode to have a published accuracy score. Start with a small, representative library and inspect the citations before trusting answers in high-stakes work.