Sometimes you do not want an answer synthesized for you. You want to search the source material directly, scan the results, and read the documents yourself. And sometimes the document you need is not part of any curated library. It is your own equipment manual, your own regional field guide, your own notes.
AI Document Search brings both capabilities to the Off-Grid AI device: semantic search across the entire verified corpus, and a personal library where you can upload your own PDFs and search them the same way. All of it runs on the device. Nothing touches the internet at any point.
This post explains how offline document search works, what happens when you upload a PDF, and the single most important design decision in the entire feature: the wall.
What Offline AI Document Search Is
Offline AI document search is semantic search that runs entirely on a local device: your query is converted to an embedding, compared against pre-indexed document content, and matched by meaning rather than exact keywords, with no network connection involved.
Keyword search finds the word "tourniquet." Semantic search finds the passage about controlling severe limb bleeding even if it never uses the word you typed. On a device built for stress conditions, searching by meaning is not a luxury. It is how you find the right page when you do not know the right terminology.
Two Modes, One Hard Wall Between Them
The Off-Grid AI device now has two distinct surfaces:
- Ask mode: the gated survival assistant. Retrieval, citation, validation, fail-closed behavior. Answers come only from the verified corpus.
- Document Search mode: direct semantic search over documents, including ones you upload yourself.
The most important property of the system is what these two modes never share. Your uploaded documents are searched only in Document Search mode. They never feed the survival assistant. If you upload a PDF containing dangerous nonsense, Ask mode will never surface it. The survival assistant answers from the verified corpus, full stop.
This is not a policy written in a document somewhere. It is enforced by an automated contamination test in the build itself: a document containing a unique nonsense token is indexed, and the build fails unless Ask mode is provably unable to find it while Document Search can. If anyone ever wires user documents into the survival path by accident, the software does not ship.
A survival tool that can be contaminated by user uploads is not a survival tool. The wall is the feature.
What Happens When You Upload a PDF
Adding your own documents is a four-step, fully local pipeline:
- 1. Validation and extraction. The device confirms the file is a real PDF and extracts its text page by page.
- 2. Honest rejection of scanned files. If a PDF is only images with no text layer, it is rejected immediately with a clear message. The device does not ship OCR, and silently indexing garbage would be worse than saying no.
- 3. Local indexing. The text goes to the device's own local server, never to any external service, where it is chunked and embedded.
- 4. A personal search index is built and stored alongside your document, with streaming progress the whole way. A 100-page PDF indexes in roughly half a minute on-device.
Links First, Summaries Second
When you run a search, results arrive in a deliberate order. Document titles and matched-page links stream in first. AI-generated summaries follow after.
The reasoning is about failure modes. Links are deterministic and fast. Summaries require the language model, can be slow, and can fail. So the reliable part arrives immediately, and the fragile part is treated as a bonus. If a summary fails, you see "Summary unavailable" and every link still works. The feature degrades gracefully instead of breaking.
The summaries themselves are deliberately constrained: generated strictly from the document's own text, forbidden from adding facts or giving advice, and clearly labeled as machine-generated and unverified.
Honest Labeling on Every Result
Trust levels are visible on every card:
- Curated library documents carry a calm "Curated library" badge.
- Your uploads carry an amber "Your document, not curated" badge.
- A standing banner steers you back to Ask mode for anything survival-critical or medical, because raw document search is the wrong tool for a life-or-death question.
The Unglamorous Safeguards
Accepting user files is where a lot of software gets hurt, so the pipeline is defensive by default: disk-space checks before any upload is accepted, size caps enforced mid-transfer, duplicate detection by hash, caps on total library size, and storage by server-generated ID so your filename is display text, never a trusted path.
None of this is exciting. All of it is the difference between a demo and a device you can stake decisions on.
Real-World Uses for a Personal Offline Library
- Homestead operations: index your generator, well pump, and solar charge controller manuals alongside the Master Electrician Pack and Master Mechanic Pack. Search them by symptom, not by page number, when something fails at night.
- Regional preparedness: add your county emergency plans and local plant or hazard guides, searchable alongside the curated corpus.
- Field professions: load equipment documentation before deploying somewhere with no connectivity, and search it on-site.
- Long-term storage of knowledge: your documents and their search index live on the device, immune to link rot, paywalls, and outages.
Common Misconceptions About Uploading Your Own Documents
"Uploading a document teaches the AI"
It does not, and that is deliberate. Your documents are indexed for search, not folded into the survival assistant's knowledge. The assistant's answers come from the verified corpus only, which is what makes them worth trusting. Your library extends what you can search, never what the assistant will assert.
"Semantic search means the AI decides what is true"
Semantic search finds relevant passages by meaning. It makes no truth judgments at all, which is why every result deep-links into the actual document so you read the primary text yourself, and why summaries are labeled machine-generated and barred from adding facts.
"Any PDF will work"
Scanned, image-only PDFs are rejected up front with a clear message, because the device does not ship OCR and pretending to index unreadable pages would produce silent garbage. If your critical documents are scans, run them through OCR software first, then upload the text-layer version.
Frequently Asked Questions
Where does my document data go during indexing?
To the device's own local server and nowhere else. Extraction, chunking, embedding, and index storage all happen on the hardware in front of you. No step in the pipeline has a network dependency.
How fast is indexing and search?
A 100-page PDF indexes in roughly 30 seconds on-device, with streaming progress. Searches return in milliseconds, because results ride on the same embedding search engine that already powers the rest of the platform.
Can uploaded documents affect medical or survival answers?
No, structurally. An automated contamination test in the build verifies the survival assistant cannot see user documents, and the build fails if that ever changes. This is enforced by software, not policy.
Is there a limit to my personal library?
Yes: per-file size caps and total library caps, checked against available disk space before any upload is accepted. The limits exist so document storage can never crowd out the device's core function.
The Bottom Line
Document Search extends the Off-Grid AI device from a question-answering system into a full offline research tool, without compromising the survival assistant by a single byte. The curated corpus stays sacred. Your documents stay yours, searchable and clearly labeled. And the boundary between them is enforced by the build system itself, not by good intentions.
This is the third post in a six-part series on the newest Off-Grid AI capabilities. Next: giving the device a third, much bigger brain, and making it fast even without a GPU.
