Part 3 of 3. 391 real SCOTUS cases, four retrieval strategies running side by side, multi-hop Cypher that no hybrid search can match, and the production-ready 3-stage architecture you should actually ship.
Part 1 of 3. Users ask three shapes of questions and only one of them needs a graph. Honest benchmarks, a 3-stage retrieval architecture, and why graph is a multiplier — not a replacement.
A layered training corpus — domain pairs, public ballast, and a reusable Postgres syntax corpus — is 80% of the work for a NL2SQL LoRA. The training config is YAML and patience.
A step-by-step walkthrough of building a PostgreSQL semantic layer in pg_agents — crawl the schema, enrich it, define your vocabulary, lock down categoricals, and promote the queries that work.
Raw LLMs hit 10-20% accuracy on real enterprise schemas with cryptic column names and tribal-knowledge joins. Here’s why, and the semantic-layer fix that takes you from toy to production.
