Each answer request carries its own query and freshly confirmed evidence. History is not a source.

Last updated: 2026-11-05 Product area: Cookbook Page type: Tutorial

Chat A and Chat B on one collection — Compare mode is two chats on one collection. Each Run sends that side's question. It does not store a memory the next turn trusts as evidence.

Repeated questions

Ask a second question after the first grounded answer, and see that the API does not receive your chat log unless you put it in query. Hands-on time is about 15 minutes.

Purpose

You can describe a follow-up without inventing memory. Level: same as Scoped search. There is no MongoDB Atlas connector and no chat-message import.

Prerequisites

You already hold one grounded result and its evidence refs. Those refs may go stale. Plan to search again.

Sample

First query: the 14-day window. Second query: Who pays for late Harbor Sample delivery? The note does not say. Expected status: insufficient_evidence. Do not let a model invent a party.

Steps

On Search and Answer, send the second question as a new question. Compare mode runs Chat A and Chat B separately. Empty Chat B reuses Chat A’s question only as a form default, not as retrieved memory.
For HTTP, send a new POST /search with the new query text. Then confirm the new refs. The previous answer body is not attached.
If your own application stores the prior answer, label that store as yours. It is not Bundata evidence. Do not place it in evidenceRefs.
If you fold the prior answer into query, the model may see your text, and citations still have to come from confirmed bites.

Fields

You send	The service treats it as
New `query`	The question for this call only.
New `evidenceRefs`	The only passages an answer may cite.
Your database of old answers	Out of band. Not a bite.
Atlas collection or chat-message JSON	Not a Bundata ingest.

Example

{"query":"Who pays for late Harbor Sample delivery?","confirmed":true,"evidenceRefs":["paste-a-current-ref"]}

Recovery

Stale refs return source_changed. Search again. If the note has no payer, accept insufficient_evidence. Do not retry until the model guesses.

Limits

Workspace search (POST /api/w/{workspaceId}/search/answer) is stateless per request unless you fold context into query, and there is no MongoDB Atlas source. In contrast, the collection Ask tab (POST /api/collections/{id}/answer) keeps private per-member Ask history and standing notes in stored workspaces, and external agents can use Governed agent memory (/api/agent/v1/memories/*). Even there, earlier answers and agent memories are context only—never evidence—and every claim must still cite a current passage.

Check

Success: the second call’s citations, if any, come from refs you confirmed for that call. Failure: a payer name with no citation is rejected by you, and the status you record is insufficient.