How Each Architecture Works
Data flow through the three options — from document to answer.
Vector DB / RAG Only
Pure semantic search with embeddings
~70% accuracy ceiling. No production legal AI system uses this alone. Embeddings fuzz out exact clause numbers, statute citations, and party names.
Hybrid Vector + BM25 + Cross-encoder Rerank
Two parallel retrieval paths fused before the LLM
The production floor. Harvey AI, Brazilian legal platform (184,895 answers), and HyPA-RAG (NAACL 2025) all use this pattern. BM25 catches exact matches (statute § numbers, party names, dates). Cross-encoder reranks for relevance. Non-negotiable baseline.
Full Hybrid with Knowledge Graph
KG pre-filtering + retrieval + hallucination verification
Adds two KG capabilities: (A) structured pre-filtering narrows candidates by metadata before vector search, and (B) post-generation entity grounding verifies citations against the KG. HalluGraph: AUC 0.94 for entity error detection vs BERTScore 0.60. But partially research-grade — build toward it, don’t start here.