Pipeline
Query → entity/intent extraction → metadata filters
→ chapter/topic routing
→ parallel lexical + vector + exact-term retrieval
→ optional graph/relationship expansion
→ fusion/reranking → parent-section expansion
→ evidence-quality validation → evidence packet
Why hierarchy
Atomic chunks find exact facts but may omit exceptions. Chapter summaries provide orientation but omit specifics. Return decisive passages plus enough parent context.
Channels
PostgreSQL full-text, pgvector/Qdrant semantic search, exact ingredient/variety/district/rule lookup, metadata filters, chapter summaries, optional knowledge graph and similar verified cases.
Iterative planning
The model may issue subqueries for symptoms, competing diagnosis, environmental fit, treatment/registration and contradictory evidence. Each query records its purpose.
Ranking
Authority, local applicability, freshness, semantic relevance, exact-term match and source diversity.
Evaluation
Recall@k, decisive-passage recall, parent-context sufficiency, authority/date correctness, citation correctness, cross-language retrieval and unsupported-answer rate.
Pilot
Start with PostgreSQL FTS + pgvector. Add Qdrant/OpenSearch only when measured quality/scale requires.