Boost scores when query grade matches standard title grade, penalize mismatches. Add part disambiguation to correctly route queries to specific standard parts (e.g., IS 12269 (Part 1) vs (Part 2)). Regenerate retrieval results with improved ranking.
Add Pass 3 to recover scope text incorrectly placed in previous block, and Pass 4 to truncate bleed from the following standard. Regenerate standards.json and standards_chunks.json with the improved parser.
- Add faiss-cpu, rank-bm25, sentence-transformers, numpy to requirements.txt.
(previously only pymupdf was listed; other deps were manual-install only)
- Cast score to float() before round() to avoid numpy type serialization errors.
- Pass expected_standards through _format_result for eval script compatibility.
- Update retrieval_results.json with expected_standards per query for eval.
Kshitij
