Every Off-Grid AI answer has always shipped with a "Sources & Reasoning" panel. That was the baseline: no uncited claims, ever. But a citation you cannot inspect is a promise, not proof. If the answer says a passage came from a specific field manual, you should be able to open that manual, on the device, with no internet, and read the passage yourself.
Now you can. Every eligible citation is a tappable link that opens the original source PDF, scrolled to the exact section the answer came from, with the passage highlighted. Hundreds of source documents ride on the device itself. Zero network required.
This post covers what it took to make citations physically inspectable on an offline AI system, and why this capability is the dividing line between AI you can trust and AI you have to take on faith.
What Citation-Backed Offline AI Actually Means
A citation-backed offline AI system is one where every factual answer is generated from retrieved source documents, and those documents are stored on the device itself, so any claim can be verified against its original source without an internet connection.
Three properties have to hold at once:
- Grounding: answers come from real documents, not model memory.
- Attribution: every claim points to a specific document and location.
- Inspectability: the source itself is available, offline, one tap away.
Most AI products offer none of these. Some offer the first two. Off-Grid AI now delivers all three.
The Engineering Problem: Finding the Exact Page, Offline
Linking a citation to a document sounds trivial. Linking it to the exact page and passage, across hundreds of PDFs, on a device with no internet, is not.
Cryptographic Provenance, Not Guessable Paths
Source records in the corpus are joined to their raw files by SHA-256 hash, never by filename or path. A citation connects to a document only if the file on the device is cryptographically identical to the one the corpus was built from. Wrong file, no link. This is what deterministic provenance looks like in practice.
Precision Page Mapping
The system maps each cited chunk of text back to its exact page using layered matching: exact string match with binary search over page offsets first, then fuzzy matching for older or differently processed documents, then coarser fallbacks. A hard quality bar governs the release: at least 90 percent of core citations must resolve to a precise page through the top matching tiers. A link that lies about where content lives is worse than no link.
A Bundled, Offline PDF Viewer
Not every operating system renders PDFs natively, and "it should just open" is not an engineering plan. The device carries its own full PDF viewer, served from its own local server, behaving identically on macOS, Windows, and Linux. The viewer jumps to the cited page, highlights the passage, and leaves the whole document scrollable, because sometimes you need to read around the answer, not just the answer.
Licensing, Handled the Right Way
A serious offline library is a mix of public-domain and licensed material, and "licensed for building the corpus" is not "licensed for redistribution." The rule here is strict and fail-safe:
- Public-domain sources ship on the device with a working document link.
- Licensed sources remain fully cited, with title, author, and year visible, but carry no bundled file.
- Unknown license status is treated as non-distributable. Unknown means no.
- Expansion pack documents, such as the Veterinary Pack library, ride along locked, and open once the corresponding pack is installed and licensed.
This is what field-ready looks like at the legal layer: no gray areas, no silent exceptions.
Fail-Soft by Design
One principle governed the whole feature: source linking can never break the ability to get an answer. If a page map is missing, if a file is absent, if a catalog entry does not resolve, the citation simply renders the way it always has. The feature is pure upside. When it works, you get the document. When it cannot, you are no worse off than before.
An offline AI system for emergencies has to degrade gracefully at every layer. Any feature that can take the core function down with it does not belong on the device.
Why Inspectable Sources Matter in the Field
Consider the situations this device is built for:
- Grid-down water treatment. The answer gives a dosage. You open the source manual and confirm the table yourself before treating water for your family.
- Remote medical decisions. A field-medicine answer cites a protocol. You read the full protocol in context, including the contraindications surrounding the cited passage.
- Equipment repair off-grid. The answer references a procedure. You open the manual to the diagram on the same page and work from the primary source.
In every case, the AI is the index, not the authority. The documents are the authority. That inversion is the core of trustworthy offline intelligence.
How This Compares to Other Approaches
- Cloud AI: may link to web sources, all of which are unreachable the moment connectivity fails.
- DIY local models: generate answers from training weights with no sources at all. Verification is impossible by construction.
- Offline document dumps: give you gigabytes of PDFs and no way to find the right page under stress.
- Off-Grid AI: retrieval-based answers, per-claim citations, and one-tap access to the exact cited page, all offline.
Common Misconceptions About AI Citations
"A citation label means the answer is verified"
Only if you can inspect the source. Plenty of AI products display source names that are themselves generated text, pointing at documents that were never actually retrieved, or that do not exist. A trustworthy citation is one that resolves to a real file, at a real page, that you can open. That is the bar this feature was built to clear.
"Storing documents locally is just an offline cache"
A cache is best-effort. This is provenance. Every shipped document is joined to its citations by cryptographic hash, so the file you open is provably the file the answer was built from. If the hashes do not match, there is no link. No guessable filenames, no near-enough versions.
"More sources always means better answers"
Quantity without licensing discipline and page-level mapping produces a pile of PDFs, not a reference system. The value is in the joinery: precise page maps, verified files, and honest handling of what can and cannot be redistributed.
Frequently Asked Questions
Do source documents take up significant space?
The distributable source library represents several gigabytes of the device's storage, and future library growth arrives as small deltas, so expanding the corpus never means re-downloading everything.
Why do some citations have no document link?
Because the source is licensed rather than public domain, or its license status could not be confirmed. Those answers remain fully cited with title, author, and year. The device simply will not distribute a file it does not have clear rights to, and unknown status is treated as no.
Does opening a source PDF require any special software?
No. The viewer is bundled on the device and served locally, so behavior is identical across macOS, Windows, and Linux, with no installs and no network.
What happens if a page map is missing for a document?
The citation renders exactly as it did before this feature existed. Source linking is engineered as pure upside: it can enhance an answer, and it can never break one.
The Bottom Line
Any AI can sound confident. Very few can hand you the page they got the answer from. Fewer still can do it on a device with no internet connection, joined by cryptographic hash, with licensing handled honestly and every failure mode designed to fall back safely.
When you evaluate an offline AI assistant for preparedness, ask one question: can I open the source? If the answer is no, you are being asked to trust a black box. Off-Grid AI was built so you never have to.
This is the second post in a six-part series on the newest Off-Grid AI capabilities. Next: turning the on-device library into a full offline document search engine, including your own PDFs.
