October 19, 2026 · QueryCrunch
The answer quality of a RAG system is mostly a retrieval problem. Chunk sizes that match how people ask, the retrievers syntax for composing BM25 and kNN, a reranker on the top 50 rather than the top 500, and a judgment list so you can tell whether any of it helped.
Read moreSeptember 21, 2026 · QueryCrunch
kNN search is fast when the vectors fit in page cache and miserable when they don't. What int8, int4, and binary quantization actually cost in recall, how oversampling and rescoring buy it back, and how to size the nodes before you commit.
Read moreSeptember 17, 2026 · QueryCrunch
Elasticsearch can now own chunking, embedding, and query-time inference behind a single field type. What semantic_text actually does, the retriever syntax that goes with it, where the latency and cost land, and the OpenSearch neural-search equivalent.
Read moreSeptember 14, 2026 · QueryCrunch
A p99 that drifted from 80 ms to 900 ms is a measurement problem first. The order we work through it on Elasticsearch and OpenSearch: took vs. client time, the slow log, the profile API, thread pool queues, caches, and the fixes that actually move the number.
Read moreSeptember 10, 2026 · QueryCrunch
Vector search does not replace BM25 — it covers a different failure mode. How to add dense retrieval to an existing Elasticsearch or OpenSearch index, fuse it with RRF, size the cost, and measure whether it actually helped.
Read moreJune 30, 2026 · QueryCrunch
Before you reach for vector search, get the fundamentals in order: a judgment list, a metric, analyzer and mapping fixes, and measured query tuning. A first-pass process that works on Elasticsearch and OpenSearch alike.
Read more