Practical write-ups from production work: retrieval, relevance, performance and the engineering it takes to ship AI that holds up.
We tested both BBQ index types on Elasticsearch 9.5.5. On self-managed clusters the default is still bbq_hnsw, bbq_disk needs an Enterprise license to index, and switching between them means a reindex.
nmslib was deprecated in k-NN 2.19 and OpenSearch 3.0 refuses to create new nmslib indices. We upgraded a 2.19.7 nmslib index to 3.9.0, found every nmslib field, reindexed to Faiss and swapped the alias. Here is each request.
TurboQuant arrived in Qdrant 1.18, and Qdrant calls 4-bit a good default for many collections. Compression per bit depth, which encodings rescore by default, how oversampling works, and a script that measures recall@10 on your data.
The first query to a cold turbopuffer namespace reads object storage and is tens of times slower than a cached one. Here is why, what the hint_cache_warm endpoint costs, and where to call it in an agent loop.
Vespa runs reciprocal rank fusion in the global-phase, over the top rerank-count hits from first-phase. We tested it on Vespa 8.778 and show how a small rerank-count drops a strong keyword match from the fusion.
When an agent gets answers wrong, teams reach for a bigger model. Published benchmarks point somewhere cheaper: the same agent with a better retriever jumped from 55.9% to 70.1% accuracy and searched less doing it.
A 20-step agent that gets each step right 95% of the time finishes cleanly only 36% of the time. Here is the arithmetic, what benchmarks show, and the five design levers that actually move reliability.
Most RAG systems that disappoint in production fail in a handful of predictable ways. For each one: the symptom, the metric that catches it, and the fix.
We ran the same hybrid queries on Elasticsearch 9.5.5 and OpenSearch 3.9.0 with identical data and vectors. Raw scores differ by exactly 2.2x, fused scores match to four decimals, and your page size can change the ranking.
Migrate to OpenSearch when Apache 2.0 licensing, cost or vendor neutrality matter more than Elastic-only features. Plan with a compatibility check, a migration method and a rollback plan.
An OpenSearch search pipeline is a chain of processors that rewrites search requests and transforms results inside OpenSearch, so query logic lives in the cluster instead of your app.
OpenSearch hybrid search runs a BM25 query and a vector query together, then a search pipeline normalizes and combines their scores (or fuses ranks with RRF) into one ranked list.
The biggest Elasticsearch performance wins come from right-sized shards (10-50GB), correct field types, filter context, bulk indexing, and leaving heap auto-sizing on.
Speed up Elasticsearch queries by moving non-scoring clauses to filter context, mapping exact-match fields as keyword, returning less data, sizing shards at 10-50GB and profiling slow queries.
Upgrade Elasticsearch without downtime with a rolling upgrade: one node at a time, data tiers first and master-eligible nodes last, waiting for green health between nodes.
Improve Elasticsearch relevance by boosting fields and signals with function_score, adding search-time synonyms, combining BM25 with vector search, and measuring every change.
Secure Elasticsearch by keeping the default security on, using TLS on both layers, least-privilege roles with field and document security, scoped API keys, and audit logging.
We write about the retrieval layer behind AI: RAG, agents, vector and hybrid search, and the engines that run them, drawn from production work.
Retrieval-augmented generation (RAG) is only as good as the search behind it. These posts cover why RAG systems fail in production, how to fix retrieval and relevance, and what separates a demo from a system you can trust.
Agents multiply small errors across steps, so reliability, tool design and retrieval quality decide whether they work. Read the math, then model your own agent with the free calculator.
Vector search finds results by meaning, and hybrid search blends it with keyword matching. These posts explain how hybrid scores are combined and ranked, with tools to size your index and test fusion yourself.
Performance tuning, security, zero-downtime upgrades and migration guides for teams running Elasticsearch or OpenSearch, the engines behind many search and RAG stacks.
