amal66 makes chat quotes earn their citations
This fork stops a polished AI answer from passing off invented document language as a real quote.
Before a chat response presents a quotation from an uploaded document, amal66's version checks it against the document itself. It does this with direct text matching, not another AI judgment, so it adds neither another model's guess nor another model call.
- Verified: the quoted words appear exactly in the source.
- Repaired: minor differences in case, spacing, or punctuation are corrected to the source's actual wording.
- Unverified: a quote that cannot be found stays visible, but is clearly not trusted.
The check also handles quotations spanning pages and avoids treating an unreadable document as evidence. A narrow punctuation edge case is documented for follow-up, rather than hidden.
Spotted something wrong? Or know the PR text has fresher detail than the writeup above?