How every answer shows its sources

Imagine a careful paralegal. Ask them a question about a contract and they never answer from memory: they answer, then slide the binder across the desk, open to the tab and page the answer came from, with the relevant passage flagged. You can accept the summary, or you can read the source yourself. Either way, the claim and its evidence travel together.

Citation-preserving RAG makes your node that paralegal. Every answer it generates carries numbered markers like [1] and [2] after the claims it makes, and each number resolves to a citation: the exact excerpt the answer drew on, the document it belongs to, the page it sits on (for paginated formats like PDF), and the precise character range within the document. Nothing about the trail is reconstructed after the fact; it is carried through from the moment the document was ingested.

The trail starts at ingestion. When your node cuts a document into excerpts, each excerpt is stamped on the spot with where it came from: the source file or URL, the document's registry identity, its position in the document, its page, and its character offsets. Those stamps live alongside the excerpt in your node's own database and ride along through hybrid search untouched.

At answer time, the excerpts that ground the answer are handed to the model as a numbered list, and the model is instructed to mark each claim with the number of the excerpt supporting it. The numbering in the answer and the numbering in the citation list are the same list in the same order, so a marker resolves to its source by position, with no fuzzy matching involved. A marker pointing at a number that was never provided is ignored rather than trusted: the model cannot invent a source and have it accepted.

For regulated work, an unattributed answer is an unusable answer: a claim you cannot trace is a claim you cannot file, audit, or defend. Citations turn a generated answer from something you must take on faith into something you can verify against your own documents, which never left your deployment in the first place. The provenance stamps are recorded where your documents are stored, and attribution is computed where retrieval runs; no third-party service is consulted to decide what grounded your answer.

A citation tells you which excerpt grounded a claim; it does not certify that the claim restates the excerpt faithfully. The model decides where to place markers, so a claim can occasionally arrive unmarked, or marked against a less relevant excerpt than you would have chosen; the full list of retrieved excerpts is always returned so you can check the answer against all of them. Documents ingested before the document registry existed, or written through the raw chunk path rather than the document pipeline, carry no source stamps: their citations degrade to the excerpt text alone. And page numbers exist only for formats that have pages; plain text and web pages cite by character position instead.