This feature is in public preview.
How routing decisions are made
When an end user asks a question, KAT:- Extracts intent and any slot values from the query.
- Looks up each knowledge base’s manifest to determine which domains can answer.
- Picks one or more knowledge bases to query based on relevance and required slots.
- Issues retrieval calls in parallel.
- Synthesizes a single answer from the results, with citations.
OUT_OF_SCOPE. If a needed slot is missing, KAT returns ASK_SLOT.
Designing knowledge bases for routing
Routing works best when knowledge bases have non-overlapping scope. Some practical guidance:- Group documents by domain, not by source. A single knowledge base can mix sources from one domain; a single source can be split across knowledge bases if needed.
- Give each knowledge base a name and description that reflects its scope.
- Re-run introspection after major content changes so the manifest stays accurate.
When to use which strategy
For what each strategy does, see KAT overview. The table below is selection guidance.Mixing knowledge bases of different sizes
Knowledge bases do not need to be the same size. KAT uses manifests, not document counts, to make routing decisions. A small knowledge base with a focused scope routes just as cleanly as a large one with broad coverage.Observability
The deployment dashboard logs every routing decision with the outcome (CALL_KB, ASK_SLOT, BLOCKED, or OUT_OF_SCOPE) and the knowledge bases queried. Use this to identify questions that consistently land on the wrong knowledge base or get refused unexpectedly. See Analytics and event logs.