Search and answers

Hybrid retrieval

Search runs two scorers over the same chunks and combines them:

• Keyword. Postgres tsvector scoring, which finds the exact identifier, error string, or product name someone typed.
• Semantic. pgvector cosine similarity over chunk embeddings, which finds the passage that means the same thing in different words.

Tag weights bias the result, so a document you marked canonical outranks a draft that happens to phrase things closer to the query.

Neither scorer alone is enough. Keyword misses paraphrase, and vectors miss the exact string you are searching for.

Answers

Ask in the agent dock, a knowledge chat destination, or the Slack bot, and generation runs on top of retrieval. The reply streams in with numbered citations, and each number is the document, commit, or thread it came from. Follow the citation before you trust the sentence. When the sources do not cover something, Mari says so in one sentence instead of guessing.

A good run can be promoted into an approved answer on Workflows. An approved answer is a reusable one, not just a transcript of a question someone asked once, and it is served verbatim next time the same question arrives. See Sync schedules and workflows.

Contradictions

Mari looks for assertions that disagree with each other across sources. That is usually two documents that were both right when written, and one of them stopped being right without anyone noticing. See Facts for what to do about it.

Honest counters

Usage counters count real events from the day counting started. They are not backfilled to look busier.