a729bc1271
Improve follow-up question quality by feeding the model structured context (evidence, capabilities, complete turns) instead of raw RAG content, and rank knowledge candidates by current-topic relevance.