Vector Database Shakeup: Postgres and Azure SQL Absorb AI Search

The vector database is becoming a feature

Over the past five weeks, a series of product announcements has pointed in one direction. Vector search, the retrieval layer behind most generative AI applications, is moving out of specialized products and into the general-purpose databases, search engines and object stores that enterprises already run. The specialist vendors are responding by changing what they sell.

The newest example came on October 13. Tiger Data, the company behind TimescaleDB, released pg_vectorsearch, an open-source PostgreSQL extension built around a new index it calls PRISM.12 The company claims a median latency of 6 milliseconds at 95% recall on 100 million vectors, measured on a machine where a pgvector HNSW index build never finished.12 Its pitch is that developers should be able to stay inside Postgres.12 Tiger Data also calls this first public release an alpha. The release notes say it is meant for testing and not for production. Automatic rebalancing is missing, space from deleted rows is not reclaimed without a REINDEX, and replication and backup testing are still pending.12 Independent commentary has pointed out that filtered queries can return fewer results than requested, and filtered "streaming search" is planned only for general availability.12

The benchmark number is not the main point. What matters is the strategy behind it: a time-series database company betting that most teams would rather extend Postgres than adopt another datastore.

Hyperscalers move vector search into general availability

The large cloud providers are making the same bet with production-grade releases. On September 28, Microsoft listed vector search and vector indexes in Azure SQL Database as generally available. That covers the CREATE VECTOR INDEX statement and the VECTOR_SEARCH function, both built on the DiskANN algorithm.15 Microsoft says customers can build AI features on enterprise data without a separate vector database, while keeping Azure SQL's existing security and governance.15

The path there was rocky. In the first public preview, a table became read-only once an index was created on it. The Azure SQL team later admitted it shipped early because demand was so strong.15 A later index format added full support for inserts, updates and deletes. It also applies filters during the vector search rather than after it, and lets the optimizer choose between DiskANN and exact search.15 Some reference documentation still carried "Preview" labels after the announcement, and GA status was confirmed only for Azure SQL Database, not Managed Instance.15 Buyers should read the GA label narrowly.

Google has taken a similar path with AlloyDB for PostgreSQL. On September 21, it made generally available a feature that uses AlloyDB's columnar engine as an in-memory, read-optimized cache for HNSW vector indexes, aimed at raising queries per second.13 That came after an April preview of "vector assist" for setting up production vector search, automatic tuning and maintenance for ScaNN indexes, and automated embedding management earlier in the year.13 On AWS, Amazon S3 Vectors added metadata pre-filtering. Filters are applied before the similarity search, which AWS says returns up to five times more matching vectors when a filter is selective.17 Oracle keeps adding AI Vector Search features through quarterly release updates to its 26ai database, most recently in July 2026.20

Filtering is the common thread. Azure moved to filtering during the search, AWS moved filtering ahead of it, and the main open question about Tiger Data's alpha is filtered recall.151712 Fast nearest-neighbor search on its own is now widely available. Getting accurate results when a query also has business filters attached is where vendors are competing.

Elastic's consolidation pitch

Elastic presented its move as a new product. The Elasticsearch Vector Database, announced on September 11, runs on Elastic Cloud Serverless and is aimed at retrieval-augmented generation and AI agent workloads at the scale of hundreds of billions of vectors.11 It ships with preset configurations for chunking, embedding and hybrid retrieval. It supports text, image and multimodal vectors in a single index, and it includes managed GPU inference through Elastic Inference Service.11 Elastic says its Better Binary Quantization and DiskBBQ techniques cut memory use by up to 32 times while keeping recall strong.11 It also prices by data volume and search capacity rather than compute units, which should make bills easier to forecast.11

The argument is the same one Microsoft and Google are making. Customers already running Elasticsearch for search and observability should not need a separate system for AI retrieval.11 The vector features underneath this launch reached general availability in August, so the September announcement mostly packaged work that was already in production.11

The specialists change what they sell

If general-purpose databases are taking over basic vector search, dedicated vendors have to offer something else. Three moves this quarter show how they are adapting.

The most striking came from Turbopuffer, a vector search vendor reported to have about $100 million in annual recurring revenue. It published a post titled "RIP, vector database" and said its v3 redesign drops approximate nearest-neighbor search as the primary way data is organized.14 One analysis treated this as the end of the vector-first design era. It also argued that pgvector with pgvectorscale is the right default for most RAG workloads, citing a benchmark of 471 queries per second at 50 million vectors and roughly 11 times Qdrant's throughput.14 That analysis is openly opinionated, and the benchmark should be treated as one data point. Still, a vector vendor declaring the category dead is a notable signal.

Zilliz, the company behind Milvus, has rebranded its cloud service from a vector database into a "Vector Lakebase." The original database becomes the real-time serving layer, sitting on top of an S3-based data foundation that also handles on-demand search and analytics.19 Zilliz justifies the change with cost: exploratory and batch workloads often run over datasets far larger than live serving data, and their compute sits idle more than 97% of the time, which makes always-on clusters hard to justify.19 Milvus 3.0 entered public preview at the same time, adding open data formats and integration with existing data lakes.19

Pinecone took a different route. It ended its single-cloud lock-in by offering bring-your-own-cloud deployments on AWS, Azure and Google Cloud. Customer vectors, metadata and request payloads stay inside the customer's own account, while Pinecone's control plane handles operations without inbound network access.18 Trade coverage noted that Zilliz already offered BYOC on all three clouds, so Pinecone is catching up here rather than leading.18

These moves have a shared logic. A standalone vector index is no longer enough to sell, so the specialists are competing on data residency, lakehouse economics and new storage designs.

Streaming infrastructure targets agents

The same pressure is changing streaming platforms. Agents need fresh context, and the streaming vendors want to supply it. Confluent's September product release raised scaling limits to 15 GB/s of ingress and 45 GB/s of egress on its largest dedicated clusters.25 The release also promoted Confluent Intelligence, a set of AI functions, streaming agents and a real-time context engine built on Kafka and Flink.25 For WarpStream, Confluent's diskless Kafka-compatible product, a feature called Lightning Topics acknowledges writes as soon as data is durably stored in object storage. Paired with S3 Express One Zone, Confluent says this cuts median produce latency to 33 milliseconds, a 70% reduction.25

Ownership is the wider context. IBM closed its acquisition of Confluent in March 2026, reportedly for about $11 billion.29 Co-founder Jay Kreps announced in August that he is stepping back, and IBM deprecated its own Event Streams products in favor of Confluent.26 Kai Waehner's Q3 landscape raises open questions about overlap with Red Hat's Strimzi work and about whether a founder-led product pace will survive. He also observes that commercial Flink adoption never grew the way the market expected.26 That last point sits awkwardly next to Confluent's continued push to put Flink at the center of its AI pitch.25 It is the clearest disagreement in this quarter's coverage. Waehner also identifies Apache Fluss, a columnar streaming storage layer built for Flink, as the newest component to watch, though he says adoption outside Alibaba is still early.26

The takeaway

Read together, these announcements show the AI data stack consolidating. Vector search is turning into a standard database capability, available in Postgres, Azure SQL, AlloyDB, Elasticsearch and S3. Specialist vendors are moving toward lakehouse storage and customer-controlled deployments. For most enterprise teams, the practical default is now to try the database they already run before adding a new one, and to test filtered recall rather than headline latency. The products that are still alpha or carry preview labels, including pg_vectorsearch and parts of Azure's documentation, are a reminder that this shift is still in progress.