How it works
How answers are made (and why they can be trusted)
When someone asks a question, the assistant does not simply look up similar text and paraphrase it. It works the question properly against the knowledge base — restricted, at every step, to documents the asker is allowed to read.
- Access first: the asker's groups, personal grants and email grants decide what is visible before anything else happens. A document the person may not read is simply not part of their knowledge base — for any question, ever.
- The question is read the way a colleague would read it — who is asking, what they work on, and any dates it mentions all count.
- If what comes back isn't enough to answer responsibly, the assistant keeps working: it rereads the question from other angles, opens full documents when an excerpt isn't enough (contracts, agreements), and gathers everything about an entity or a period for timeline and “all of the …” questions.
- It writes the answer with numbered citations [1], [2] … — each number pointing at a specific passage it was actually shown.
- Every claim is checked against its citation before the answer is shown. A claim that cannot be backed does not survive; if nothing survives, the user gets an honest “I don't have that in the knowledge base” instead of a confident guess.
Rules the assistant answers by
- Evidence has rank: a signed instrument beats a report, a final report beats meeting notes, and everything beats drafts and stale copies. Between comparable documents, the newer governs.
- Dates are computed, not guessed: deadlines, renewal windows, “the latest version” and “as of March” resolve against the documents' own dates, exactly.
- Premises get corrected: if the question asserts something the documents contradict, the answer says so explicitly.
- It answers in the asker's language, whatever language the documents are in.
What users see
Under each answer: the numbered sources. Clicking one opens the exact passage, so anyone can verify any claim in seconds. Follow-up questions continue the same conversation — the assistant remembers the thread and re-checks it against the documents rather than trusting its own memory.
A refusal is a feature, not a failure. It means: the documents this person can read do not contain the answer. Either the knowledge isn't ingested yet, or the person lacks access — the “Check a person's view” tools under Users show which.
Last updated 29 Aug 2026