I think it is outputting the exact chunk in the embedding vs using the knowledge vector database to come up with and output the best answer.
Hoping this might work:
Enhanced _prepare_advice_from_chunks in server.py: it now keeps top_chunks, adds a use_llm flag, and (when multiple chunks are retrieved) calls a new llm_synthesizer.summarize(query, raw_candidates) helper to synthesize a coherent answer across the best chunk candidates instead of returning a single raw snippet.