Developer tests 13 RAG techniques on 46,000 chunks; only four meaningfully improved retrieval

A software developer built a retrieval-augmented generation (RAG) system over 62 ancient-history books, totalling roughly 46,000 chunks, and evaluated 13 popular retrieval techniques against a fixed 161-question test set. Starting from a baseline recall@5 of 35.2%, the experiment found that upgrading the embedding model and adding contextual chunk notes each delivered gains of around 18 points, while most other widely recommended techniques showed little or no improvement. Hybrid BM25 search produced zero gains over dense retrieval alone, and HyDE actually hurt performance by nearly 10 points. The author argues that techniques operating at the same pipeline stage tend to compete rather than compound, making careful measurement on one's own corpus essential. A key practical finding was that ingest-time improvements offer better long-term value than query-time techniques, which add latency and token costs on every request.