Vector databases were the hottest data infrastructure category of 2023. The funding got out of hand, the marketing got worse, and a lot of teams ended up with vector stores they barely used. Three years on, the market has done what these markets always do: it has consolidated, the standalone-vector-database thesis has lost ground, and the actual production patterns are different from what anyone was pitching in 2023.

The dominant 2026 pattern is hybrid search inside an existing operational database rather than a separate dedicated vector store. Postgres with pgvector keeps eating market share. Elasticsearch and OpenSearch have closed most of the functional gap. The major cloud databases all ship vector indexing now, and the ergonomics of doing vector search next to your transactional data have turned out to matter more than peak vector throughput.

The standalone vector database vendors haven’t disappeared, but their positioning has shifted. The market they’re now winning is the high-scale, low-latency, separate-from-the-database use case: search-heavy applications, recommendations at scale, and the inference layer for large agentic systems. For most enterprise RAG deployments, the calculus has gone the other way. Teams want vectors next to their structured data, governed under the same security model, with the same ops tooling.

The other thing that has shifted is what people use vectors for. The 2023 vision was that vectors would replace traditional keyword search for almost everything. The 2026 reality is that hybrid retrieval — combining lexical search, vector search, and structured filtering — beats pure vector retrieval in most enterprise search applications. The teams that ran A/B tests on real user queries reached this conclusion early. The teams that listened to vendor marketing took longer.

For production RAG, the better practitioners have moved past naive embedding-and-search. The patterns that work in 2026 use chunking strategies tuned to the specific document domain, query rewriting, hybrid retrieval with rerankers, and retrieval evaluation that runs on actual user query distributions rather than synthetic benchmarks. The infrastructure choice — pgvector, OpenSearch, dedicated vector store — is less consequential than the retrieval engineering on top.

The lesson for teams thinking about vector infrastructure in 2026 is to start with the simplest viable option that fits inside their existing data stack. For most teams, that’s pgvector or whatever vector capability their cloud database already ships. Move to a dedicated vector store only if the workload genuinely requires it: latency profile, scale, or specific feature need. The default-to-dedicated thinking from 2023 produced a lot of wasted infrastructure spend.

The vendors most likely to thrive from here are the ones that have shifted from “we are the vector store” to “we are the search and retrieval layer for AI applications.” Pure vector storage is increasingly a commodity. The retrieval engineering, the rerankers, the evaluation tooling, and the AI-native query interfaces are where the differentiation is heading.

For Australian enterprises planning AI infrastructure work in 2026, an AI consultancy with current production RAG experience is worth bringing in early. The cost of getting the retrieval architecture wrong is high, and the cost of correcting it later is higher. The decisions made in the first quarter of an AI infrastructure project tend to determine the next two years of operating cost.