RAG for appraisers — building a knowledge base from 5,000 historical cases
From chunking to reranking — a complete guide to a Hybrid RAG architecture designed for hard technical-legal documentation.
Why hybrid
Pure embeddings lose legal terminology and technical codes. Pure BM25 struggles with semantic questions. A 0.6 vector + 0.4 BM25 hybrid produces the best results across 1,200 control queries in our tests.
Chunking
Section-level 800–1,000 tokens with 100–150 overlap. Headings and metadata preserved (document_id, section, page_range, case_id).
Heuristic reranking
Boost for fragments from the same industry (automotive/machinery/marine) and the same decade of authorship.
This is an editorial skeleton — the full expert article will follow shortly. The structure, headings and SEO metadata are production-ready.