Ask the wiki ONE question and get the answer — the platform’s own agents’ read path. The recall assistant continues the caller’s threadId across questions (pages already read stay in context) and answers concurrent questions on one thread in a single run. mode: NAIVE answers in ONE model call with the venue’s whole wiki preloaded instead. Returns the answer plus an outcome (answered / no_knowledge / conflict / degraded).
The question, in natural language, exactly as asked. Bundle EVERYTHING you need answered into this one question (numbered parts: 1) … 2) … 3) …) — the assistant answers every part in one reply, each with its own page paths. One call per reply, not one call per fact.
The caller’s conversation thread (jonna passes its own thread id). The recall assistant continues that thread across questions — pages read for an earlier question stay in context — and concurrent questions on one thread are answered by ONE run. Omit for a one-off question on a fresh thread.
Bind the question to one bucket: LOCATION (a venue) or GROUP (a wiki_group, no location). Omit to fall back to locationId.
Convenience venue bind — equivalent to scope: { scopeType: LOCATION, scopeId }.
Staff-only mode: with NO binding, span every venue the caller can see and answer PER VENUE (“Alex Grill: …; Hell’s Kitchen: no information”) — the “which venue does X?” question. A binding plus crossScope keeps the binding.
The answer, in plain prose ending with the source page path(s) — part by part when the question bundled several parts (a part may say the wiki holds no information about it while the others are answered). Read it literally: a no_knowledge outcome is authoritative for the whole question (the venue has NO documented answer — do not search harder); a conflict outcome means the wiki holds two values and the fact is unresolved (never pick one); a degraded outcome means the assistant was unavailable and answer is a [wiki_recall: DEGRADED …]-stamped block of RAW search hits, not an answer.
answered | no_knowledge | conflict | degraded — see answer.
The recall run that answered — one run may answer several questions.
How many questions that run answered in one go (concurrent questions on one thread coalesce).
True when the thread already held earlier questions — the assistant reused what it had read.
Which recall path answered: agent (the ReAct assistant with wiki tools) or naive (one model call with the venue’s whole wiki preloaded, no tools).
Which recall path answers. AGENT: the ReAct assistant (wiki tools, thread continuity, coalescing). NAIVE: one model call with the venue’s whole wiki preloaded — no tools, lowest latency, cached per venue. Omit for the deployment default (RECALL_MODE). Cross-scope questions always take the agent path.
