Pincites

A pincite is a citation to the exact page or line, the way a legal brief cites a page of a judgment. Vectorless attaches one to each claim in an answer from POST /v1/answer.

How a claim gets its pincite

1. The answer model writes over the evidence pages and marks each claim [n].
2. Code lists candidate quotes from the cited page's own sentences and table rows, and keeps a shortlist of 10.
3. Jev picks one quote per claim, in one batched call for the whole answer. It can pick none.
4. Below 0.35 confidence, the citation falls back to page level.
5. The quote is located on the page exactly, or at 80% or more token overlap. Quotes under 4 tokens must match exactly.
6. The location is returned as one rectangle per line, in page coordinates normalised to 0–1, ready to draw over the page image (GET /v1/documents/{id}/pages/{n}/image).

What it guarantees, and what it does not

• No generated text can reach a citation. The quote is always text taken from the page, chosen by Jev from candidates code found. The answer model never writes a quote.
• Precision is line, page or section. There is no table-cell precision.
• It is measured on one document. On the 3M 2022 10-K (252 pages) the pincite step took 0.4–0.6 s inside /v1/answer, and every claim checked by eye landed at line precision. That is a small visual check, not a benchmark.
• A pincite shows where a claim's support is, not that the claim is right. The answer model can still misread a correct quote.

Without a TypeSafe key, quotes are picked by word overlap instead of by Jev.