Connected context in answers

The context planner selects candidates from each permitted source, ranks them by relevance, recency and entity match, and fills a bounded budget. Connection data and memory are kept distinguishable so the answer can cite evidence separately from stored context.

1. 1State the workspace when the answer must come from a specific one.
2. 2Name the service when a term could match several connected sources.
3. 3Say explicitly when you want only memory, or only live data — "based on stored workspace context" is a valid instruction.
4. 4Start a new conversation when the previous thread's subject would distort the result.

• Answer cites the wrong account → check the connected account label.
• Answer reflects another project → check workspace, then start a new conversation.
• Answer ignores a fact you stored → it may be conversation-scoped, or stored in another workspace.
• Answer mixes two topics → the conversation context carried over; reset with a new conversation.

You do not have to guess which context was used. Ask directly and the product will tell you:

"Which workspace, account and sources did you use for that answer?"
"What was missing that would make this more reliable?"

That one question resolves most "the answer feels off" reports, and it is faster than rephrasing and guessing.

Context assembly never crosses a workspace boundary, never includes another user's data, and never incorporates provider credentials. If an answer appears to use data outside the active workspace, stop and report it as a serious fault.