Tuning Semantic Search on JFMM.net – Joint Fleet Maintenance Manual
Summary
The article narrates a practitioner’s journey tuning semantic search for a massive manual (JFMM.net), detailing a shift from a Postgres-based vector store to SQLite with vector extension, the use of quantized embeddings, and a reranker to improve relevance. It also tackles pagination challenges and advocates a REST/HATEOAS approach to maintain state across pages, highlighting cost, latency, and deployment tradeoffs.