HNSW (Hierarchical Navigable Small World) is a graph-based algorithm for approximate nearest-neighbor search, used to index vector embeddings in databases.
Tiger Data says its pg_vectorsearch alpha answers at a 6 ms median and 95% recall on 100M vectors in its own tests. Its release notes say filtered queries can come back short of LIMIT, so we would keep pgvector HNSW in place for now.
Reality
Turbopuffer is rewriting its serverless database as v3, moving the ANN index out of the core of storage and making it one of several secondary indexes. A dev.to explainer of the September 30 post says the vector-first layout held back GROUP BY and aggregation queries.
Reality
Amazon OpenSearch Service needs about 1.3 TB of resident RAM for 100 million 1,536-dimension FP32 vectors with one replica, a dev.to sizing post calculates. Raw vector values are about 98% of each entry, so the encoding picked before ingestion decides most of that memory and the node count behind it.
Reality
Open-source benchmark grades pre-filtering, post-filtering and Qdrant's in-index filtering on 120,000 mandi listings against exact brute-force neighbours. Filter placement is the only variable it changes, but the supplied write-up ends before any recall or latency figures.
Reality
Databricks made Lakebase Search generally available, pairing BM25 with a Postgres vector index it says runs 4 times cheaper than pgvector at 100M vectors. The vector index lives in object storage behind a cache, so large agent corpora no longer need RAM sized to the whole index.
Reality
Password-reset queries made up 38% of one support RAG bot's vector lookups, says a dev.to write-up that moves retrieval behind an HTTP cache. Any saving depends on identical queries repeating within one tenant inside five minutes.
Reality
A dev.to walkthrough of a permission-aware Postgres project puts the ACL test in a CTE that the vector ranking reads from, so the nearest-neighbour search only ever orders rows the caller may see.
Reality
A dev.to post argues that moving embeddings out of Postgres buys a second source of truth, a four-leg query path and 50 to 150 ms of internet latency. Its 8 ms pgvector counter-figure has no benchmark behind it.
Reality
A September 2026 comparison pins the workload at 10 million 1536-dimension vectors and reports Qdrant between $250 and $947 a month, while Pinecone's serverless bill is set by the gigabytes each query scans.
Reality
AWS's comparison of the three customer-managed backends credits OpenSearch with in-memory speed and S3 Vectors with sub-second queries at up to 90 percent lower storage cost. The Aurora spec sets a ceiling your embedding model has to fit under.
Reality
The paper reports 2 to 1,000 times the throughput of prior filtered-search methods at fixed recall, and says an existing HNSW library is enough to implement it. Both claims turn on how the graph is built.
Reality
A three-engine evaluation of filtered vector search on a purpose-built relational dataset credits Milvus's hybrid approximate/exact execution with stable recall, and puts pgvector's plan choice ahead of its index choice as the thing that loses it.
Reality
A dev.to design guide moves JSON documents and vector search into one PostgreSQL instance, and the evidence it offers is operational: deleted ETL pipelines and sync bugs, with no query timings on either side.
Reality
A dev.to walkthrough builds the whole retrieval path on pgvector and the azure_ai extension inside Azure Database for PostgreSQL. The costs land in the DDL, the extension allowlist, and a transaction that waits on an HTTP call.
Reality
pgvector ships hnsw.ef_search at 40, the size of the candidate list its graph walk keeps in flight, and a top-20 query with a tenant filter can come back with two rows and no error to explain it.
Reality
The dedicated store won every unfiltered benchmark this team ran, which happened to be the one query their product never issued. What they were paying for was a copy of derived data that could disagree with its source in silence.
Reality
An independent rebuild of the SQLite side says the gap is real, but the published tables measure two different queries on two different graphs, so the depth at which SQLite's plan collapses on your data is a number you have to take yourself.
Reality
One engineer's trading app hit a three-second P99 tail while average latency looked healthy. The cause sat in Postgres and the feature lookups, not the model, and cache tuning could not reach it.
Reality
HNSW indexing runs inside the SQL engine, so there is no second service and no network hop. There is also no index persistence unless you turn on an experimental flag.
Reality