Tune lexical retrieval without treating scores as confidence
Scores rank matches
Evidence Studio uses local BM25-style term scoring plus configurable title and exact-phrase contributions. A larger score indicates a stronger lexical match under that corpus and configuration. It is not a correctness probability. Corpus edits can change scores even for an unchanged question, so review ranking and expected-source cases rather than presenting a score as confidence.
Inspect chunk boundaries
Defaults are 500 characters per chunk with 60 characters of overlap. Small chunks isolate steps but can lose prerequisites; larger chunks preserve context but can introduce unrelated instructions. Test a procedure spanning a boundary. Read the returned excerpt and verify that conditions and actions appear together. Source offsets make these splits inspectable.
Use a controlled vocabulary
If customers say broadcaster while the procedure says encoder, add broadcaster = encoder as a synonym. Rules use literal normalized substring matching, so a broad trigger can expand unintended queries. Test a positive and a misleading negative example for every important rule. This is explicit vocabulary management, not automatic semantic or multilingual understanding.
Tune against unsupported questions
The default threshold is 0.1 and topK is three. A higher threshold suppresses weak matches but can reject valid short questions. A lower value can return irrelevant context. Include sparse supported questions and unsupported questions containing familiar product terms. Categories match exact names and are retrieval scope, not a document authorization boundary. Retest before publishing any change.
Product and guides
Explore creekX Agents · Quick start · Documentation. Written for v0.10 early access, 23 October 2026.
Frequently asked questions
Where should I start?
Read the separate Agents quick start and evaluate v0.10 in a private local workspace before a customer pilot.