Follow-up questions are a core failure mode for naive RAG. We added conversation context handling so retrieval can account for what the user already said.
The system still cannot assume missing facts. The purpose is to preserve state that exists, not to fabricate state that was never established.
Follow-up behavior we targeted
- "What about that?" where the referent exists in the prior turn.
- "What was her role?" where a named person was already introduced.
- "Compare it with the earlier option" where the comparison target is in the conversation state.
- "Do the same for the second source" where source ordering matters.
Boundary
Conversation context can help interpret a query. It is not automatically evidence. The runtime still needs retrieved or remembered context that can be inspected.
Failure mode
If a referent cannot be resolved, the correct behavior is to ask for clarification or surface uncertainty. Guessing creates a false grounding trail.