Skip to main content

Natural-language search

The global search bar at the top of every page accepts plain-English queries. Behind the scenes a Haiku-tier AI call parses the intent, picks the target entity (test runs, requirements, logs, evidence packs, standards), extracts filters, and returns a structured answer with deterministic click-through URLs.

What it can answer

Examples that work today: The LLM picks the target entity and the filter values. The platform builds the URL itself — the LLM never invents a link. If the platform can’t be confident, it falls back to traditional text search and surfaces the AI-suggested URLs as secondary options.

What you get back

The follow_up_questions are one-click — clicking them re-runs q with the new phrasing.

Endpoint

Returns 402 if the org is out of ai_tokens. Returns 200 + is_ai_query: false for queries the platform decided to handle with traditional search (short queries, exact-ID lookups).

Plan gating

Costs 1 ai-token per parse call (SEARCH_QUERY_PARSE task type — Haiku at 2k input). Free plan’s 100k grant is plenty.

What it does NOT do

  • It does not answer the question itself. It parses intent into a filter and links you to the structured view. The answer comes from the actual matrix / list page the link opens.
  • It does not read your repo source or your private logs to “answer questions” about them. It only uses the parsed query intent.
  • It does not invent IDs. If you ask “show me REQ-9999” and REQ-9999 doesn’t exist, the structured page will render empty — the AI doesn’t hallucinate rows into existence.

Next

AI overview

Every AI surface on the platform and what they cost.

Traceability matrix

Where most of these queries actually land.