1.Start with a retrieval baseline:

- TF-IDF or BM25 over chunk_text

2.Generate answers:

- Extractive: pick the most relevant sentence from the top chunk

- Abstractive: summarize top chunks (ensure grounding)

3.Choose citations:

- Start with top-2 retrieved chunks

- Improve precision by reranking or filtering

4.Validate locally using train.csv:

- Track answer score and citation precision separately

- Avoid over-citing