How answers are grounded, cited and scored.
Answers are composed only from retrieved passages. The agent is instructed — and evaluated — not to add facts that are not in the citations. Where sources conflict, the answer says so and cites both.
Each citation carries the document, source system, a deep link and an excerpt. Where the source supports it, links open at the exact location (page in a PDF, block in Notion, message in Slack, cell range in a spreadsheet).
A score from 0 to 1 estimating how fully the citations support the answer.
| Range | Behaviour |
|---|---|
| ≥ 0.8 | Answer given plainly |
| 0.6 – 0.8 | Answer with a caveat about what was and wasn't found |
| 0.4 – 0.6 | Partial answer, explicit gaps, suggestion of where to look |
| < 0.4 | Declines to answer; lists what was found and suggests who to ask |
Thresholds are adjustable per knowledge base under Settings → Answering.
/brewmycode sources (Slack), Sources considered (web app) or GET /queries/{id}/retrieval shows every passage retrieved and its score, not only those cited — useful for debugging a knowledge base.
Users can flag answers. Flags reach knowledge base editors and are added to the knowledge base's evaluation set, which runs on every index change and reports retrieval and answer quality over time.