Chunk for meaning

Splitting documents by token count destroys context. We chunk by structure — headings, tables, clauses — and attach metadata so retrieval can filter before it ranks.

Hybrid search and reranking

Combining keyword and vector search captures both exact terms and semantic intent. A lightweight reranker then promotes the passages most likely to answer the question.

Test with real questions

We build evaluation sets from actual user queries, tracking recall, grounding and answer quality over time. What gets measured gets better.