Knowledge Chat Lab

Run a frozen question pack against four comparable architectures: current RAG, prefetched Offer/FAQ understanding, model-directed Knowledge Tool access, and prefetched understanding plus focused RAG evidence. Reliability gates decide first; tokens and latency decide only between equally reliable candidates. Real governed actions and ephemeral agent state remain active. Build or refresh artifacts in Knowledge Browser Lab.

Knowledge mode

RAG onlyCurrent lab baseline: retrieve a small relevant set from published resources. Prefetched contextDiagnostic only: inject the bounded Offer/FAQ drafts before reasoning; automatic RAG is disabled. Agent Knowledge ToolThe HEA decides if and how to read catalog, FAQ, or source knowledge. No automatic retrieval. Prefetched + focused RAG CANDIDATEInject broad Offer/FAQ understanding, then retrieve focused source evidence for the current question.

Selected HEA and artifacts

Waiting for HEA picker…

Offers: —FAQs: —

Prefetched diagnostic focus (optional)

Available in both prefetched modes. It simulates an explicit visitor choice; the Agent Knowledge Tool must resolve selections itself.

Conversation

Choose a mode and ask a question. Change modes to reset the comparison history.

Redacted lab trace

No turn yet.

Knowledge mode

Selected HEA and artifacts

Ask the HEA

Conversation