Query rewriting, chunk size, re-ranking and metadata filters

Try 4 changes one at a time and re-run the evaluation: (1) query rewriting — let the LLM turn the user's question into 2–3 clearer search queries and merge the results (multi-query); (2) chunk size 300 vs 500 vs 800 tokens; (3) retrieve 20, then re-rank to the best 5 with a cross-encoder (as in the NLP FAQ project); (4) metadata filter, e.g. search only the HR policy when the question is about leave. Keep what helps.