development

Best Vector Databases

Sixteen vector databases compared on price, scale, licence and measured recall — the dedicated engines against the vector search already built into Postgres, Redis, Elasticsearch and MongoDB, with every published rate card and an honest account of which benchmarks survive scrutiny.

Updated August 2026

Quick answer: The best dedicated vector database for most teams in 2026 is Qdrant — Apache 2.0, written in Rust, genuinely simple to self-host, with a free-forever managed tier and a $50 million Series B raised on 12 March 2026 that removes most of the vendor-risk objection. But most teams should not buy a vector database at all yet: if your data already lives in PostgreSQL, pgvector 0.8.2 handles millions of vectors inside the database you already back up, monitor and staff, and every serious Postgres search extension in 2026 still delegates its vector work to it. Reach for a specialist when one of four things is true — you need fully managed zero-ops at scale (Pinecone, $50/month minimum on Standard), billions of vectors over a data lake (Milvus 3.0 and Zilliz Cloud), hybrid keyword-plus-vector search as a first-class query (Weaviate), or object-storage economics across enormous multi-tenant workloads (turbopuffer, from $16/month). The one caveat that matters more than any ranking: almost every performance number published in this category is vendor-run, and the one high-profile 2026 benchmark a vendor tried to reproduce was then retracted by that vendor — Elastic’s own follow-up on 13 July 2026 found the disks it was supposedly measuring were never read at all.

The current state of vector databases: August 2026

Three things changed the shape of this market between 2024 and today, and all three push in the same direction — away from buying a separate database.

First, quantisation stopped being a trade-off. Binary and low-bit quantisation now cut vector memory by 16 to 32 times at recall most applications cannot distinguish from full precision. Elasticsearch’s Better Binary Quantization (BBQ) has been the default for dense vectors of 384 dimensions and above since version 9.1, and Elastic’s testing across ten BEIR datasets found BBQ beat uncompressed float32 HNSW on NDCG@10 on nine of them. MongoDB’s automatic binary quantisation, generally available since 29 April 2025, cuts search-node RAM by 24 times. Qdrant’s 1.19.0 release on 5 August 2026 made 4-bit TurboQuant a primary storage datatype rather than an add-on. The practical effect: the RAM bill that used to justify a purpose-built vector database has largely evaporated.

Second, the incumbents shipped. PostgreSQL, Redis, Elasticsearch, OpenSearch and MongoDB all now have production vector search with quantisation, hybrid ranking and filtering. Amazon went further and made object storage itself queryable — Amazon S3 Vectors reached general availability on 2 December 2025 and now runs in 34 Regions, storing vectors at $0.06 per GB-month with a claimed “up to 90%” total cost reduction against specialised vector databases. If retrieval is a feature of your product rather than the product, the incumbent option is usually the correct one.

Third, the specialists moved up-market and sideways. Milvus 3.0, announced in July 2026, repositioned as “lake-native”, indexing vectors in place over Lance, Iceberg, Parquet and Vortex rather than copying data in. Pinecone shipped Nexus on 6 August 2026 as an agent knowledge engine rather than an index. Weaviate shipped Engram, a managed agent-memory service, in June 2026. LanceDB now describes itself as a “multimodal lakehouse”. The pure vector index is being absorbed into something larger in every case, because the index alone is no longer defensible.

The market consolidated accordingly. MyScale shut down on 28 May 2025. SingleStore was taken private by Vector Capital in a deal that closed on 16 October 2025, reported at roughly $500 million against a prior $1.3 billion valuation, though the price was never officially disclosed. Marqo pivoted to e-commerce search, launching its Sibbi agent on 7 May 2026. Vald has been dormant since its v1.7.17 release on 4 July 2025. And notably, no vector database company was acquired by a hyperscaler in 2025 or 2026 — the database M&A of that period went to Postgres instead, with Databricks acquiring Neon in May 2025 (reported at roughly $1 billion) and Snowflake acquiring Crunchy Data in June 2025 (reported at roughly $250 million). That is the clearest signal in the category about where the strategic buyers think retrieval belongs.

The four ways to buy vector search

Choose the layer before you choose the vendor. Most of the bad decisions in this category are layer errors, not vendor errors.

The economics differ sharply by layer. Extensions cost you the compute you already pay for. Dedicated self-hosted costs you engineering time. Managed dedicated vector databases cost a floor of $16 to $500 per month before you store a single vector, though usage-billed services such as Chroma Cloud and Amazon S3 Vectors have no floor at all and enterprise contracts run into five figures. Object-storage-native trades single-digit-millisecond warm latency for storage that costs cents per gigabyte and compute you only pay for when queries arrive.

Best vector databases ranked (August 2026)

Ranked by suitability for a team choosing today, weighting operational simplicity, cost at realistic scale, licence risk, filtering and hybrid-search quality, and how much of the vendor’s own performance claim survives independent scrutiny. Managed entry price is the lowest recurring commitment for a production-grade deployment; free tiers are listed separately in the pricing table below.

Three systems are deliberately excluded from the ranking. Faiss and hnswlib are libraries, not databases — they have no persistence, no filtering layer and no server, and several of the systems above embed them. Valkey, the Linux Foundation fork of Redis, reached 9.1 general availability on 19 May 2026 with Valkey Search 1.2 contributed by Google, and is the licence-clean route to Redis-style vector search, but its search module has less production history than the systems ranked here.

The dedicated vector databases in depth

1. Qdrant — best all-round vector database

Licence: Apache 2.0 · Language: Rust · Current version: 1.19.0, released 5 August 2026 · Managed free tier: free forever, 0.5 vCPU / 1 GB RAM / 4 GB disk

Qdrant is the pick for teams that want a real vector database without a real vendor relationship. It is Apache 2.0 with no open-core carve-out, ships as a single Rust binary that is the least painful of the dedicated engines to self-host, and its filtering implementation — the thing that separates usable vector search from demo vector search — is the strongest in the open-source tier.

Version 1.19.0 landed on 5 August 2026 with three changes that matter operationally: TurboQuant 4-bit as a primary storage datatype rather than a bolt-on compression layer, memory tiers that let you mark collections cold, cached or pinned, and read-affinity routing across replicas. The managed free tier is free forever rather than trial-limited — a single node with 0.5 vCPU, 1 GB RAM and 4 GB disk, which holds roughly one million 768-dimension vectors, plus free cloud inference on selected models. Standard tier is usage-based on compute, memory, storage and backup, billed hourly with a 99.5% uptime SLA. Premium adds SSO, private VPC links, 24/7 support and a 99.9% SLA rising to 99.95% on multi-availability-zone setups. Hybrid Cloud runs Qdrant’s control plane over your own infrastructure; Qdrant Serverless is listed as coming soon and is not yet available.

Qdrant closed a $50 million Series B on 12 March 2026 led by Advance Venture Partners, with Bosch Ventures, Unusual, Spark Capital and 42CAP participating, taking reported total funding to roughly $87.8 million. Compliance covers SOC 2, GDPR and HIPAA.

Why it wins: the best combination in the category of permissive licence, genuine self-hostability, strong metadata filtering, current quantisation and a funded company behind it. Nothing else scores well on all five.

Limitations: it is not the fastest system at every point on the recall curve, and Zilliz’s own VDBBench leaderboard measures Qdrant Cloud at 1,242 queries per second on a 1-million-vector Cohere dataset at a $1,000/month cost budget, well behind Zilliz’s own 13,316 — though Qdrant posts the highest recall in that table at 0.9474, and the benchmark is run by a competitor. There is no serverless option yet, so idle clusters still cost money.

Best for: teams that want one dedicated vector database, want the option to move it in-house, and are not operating at billion-vector scale.

2. Pinecone — best fully managed

Licence: proprietary · Plans: Starter $0, Builder $20/month, Standard $50/month minimum usage, Enterprise $500/month minimum usage

Pinecone remains the default answer for teams who want retrieval to be somebody else’s operational problem. There is no cluster to size, no index to rebuild and no HNSW parameter to tune, and for a large number of teams that is worth more than any benchmark result.

The pricing is now four tiers. Starter is free and includes up to 2 GB storage, 2 million write units and 1 million read units per month, 5 indexes and AWS us-east-1 only. Builder, added on 4 May 2026, is a flat $20/month where usage beyond the plan limits is blocked rather than billed — the cleanest way to cap spend in the whole category. Standard carries a $50/month minimum usage commitment; Enterprise carries $500/month and adds a 99.95% uptime SLA, BYOC deployment in your own cloud account, private endpoints, CMEK and audit logs.

The serverless unit economics Pinecone publishes officially are $0.33 per GB per month for storage, $16 per million read units in AWS us-east-1, and $0.25 per GB for bulk import from object storage. The write-unit rate is data not available from an official source — Pinecone renders it in a client-side widget, and the third-party figures in circulation ($4 and $4.50 per million write units) conflict with each other and are not vendor-published. The metering rules are worth understanding before you model cost: a query costs 1 read unit per 1 GB of namespace size with a 0.25 RU minimum, and top_k does not change the price, so the lever on query cost is namespace size, not result count. Writes cost 1 write unit per KB with a 5 WU minimum per request, which makes single-record upserts expensive relative to batches.

Pinecone publishes its own 10-million-record test — 1,024-dimension vectors, 48.8 GB — at p50 29 ms, p90 44 ms, p95 81 ms and p99 350 ms at 10 queries per second, for a total of $90.37 including import. That is a rare thing in this category: a vendor publishing its own tail latency rather than just its median.

The 2026 story is Pinecone Nexus, which reached general availability on 6 August 2026 as a knowledge engine for agents with a declarative query language called KnowQL, running its data plane inside your own cloud account. Compliance covers SOC 2 Type II, HIPAA, GDPR-readiness and ISO 27001; HIPAA is a $190/month add-on on Standard and included on Enterprise.

Why it wins: genuinely zero-operations, the most predictable spend cap in the category via the Builder plan, and the broadest compliance coverage of any dedicated vector database.

Limitations: cost at scale is the standard complaint, and Zilliz’s cost leaderboard measures Pinecone Serverless at 4.6 queries per second with a 4.85-second P99 on a 100-million-vector LAION dataset, and a $1,180 write cost to ingest it against $9.12 for Zilliz — figures that come from a direct competitor and should be read as such, but which no Pinecone-published counter-benchmark refutes. There is real corporate uncertainty too: Ash Ashutosh replaced founder Edo Liberty as CEO on 8 September 2025, the last funding round remains the $100 million Series B of April 2023 at a $750 million valuation, and Calcalist reported on 31 August 2025 that Pinecone had held early banker discussions about a sale. No acquisition has occurred as of 19 August 2026 — that remains a rumour, not an event — but it belongs in a procurement conversation. Legacy pod-based indexes have been unavailable to new customers since August 2025.

Best for: teams with no platform engineer to spare, and anyone who needs HIPAA or ISO 27001 coverage without building it.

3. Milvus and Zilliz Cloud — best at billion scale

Licence: Apache 2.0 · Current version: Milvus 3.0.0 (July 2026); maintained 2.6 line at v2.6.22, 4 August 2026 · Governance: graduated LF AI & Data project

Milvus is the open-source vector database built for the scale nobody else designs for. Milvus carries 45,600 GitHub stars, and Zilliz claims over 100 million Docker pulls and more than 10,000 enterprise deployments. Milvus’s distributed architecture separates storage, indexing and query nodes so a single deployment can hold billions of vectors.

Milvus 3.0 is the significant 2026 release, repositioning the system as “lake-native”. Its External Collections index vector, full-text, JSON and scalar data in place over Lance, Iceberg, Parquet and Vortex without copying it, with incremental sync; a new manifest-based storage engine called Loon underpins it, with Vortex as the default columnar format. It also adds point-in-time Snapshots, a Spark DataSource V2 connector for Databricks and EMR pipelines, and StructList multi-vector fields supporting late-interaction models such as ColBERT and ColPali. Note a genuine sourcing problem: Zilliz’s own announcement carries a 16 July 2026 dateline but a 1 August 2026 byline, and the Business Wire release is dated 31 July 2026 — the exact release date is disputed across Zilliz’s own materials. Python, Go and Node SDKs shipped first, with Java following.

Zilliz Cloud is the managed service. The free tier gives 5 GB storage and 2.5 million vCUs per month on Google Cloud only. Standard starts at $0 for serverless and from $126 per GB per month for dedicated clusters; Enterprise starts at $197 per month with a 99.95% uptime SLA, SAML 2.0, RBAC and private endpoints; Business Critical adds CMEK, HIPAA eligibility and a 99.99% SLA with multi-replica enabled. Zilliz’s own published cluster shapes make the storage-versus-speed trade explicit at 768 dimensions: performance-optimised holds 2 million vectors per compute unit at 500–1,500 QPS and 10 ms, from $63 per million vectors per month; capacity-optimised holds 8 million per CU at 100–300 QPS and 50–100 ms, from $16; tiered storage holds 40 million per CU at 10–50 QPS and 100–1,000 ms, from $5. Dedicated storage is $0.025 per GB per month.

Why it wins: nothing else in open source handles billions of vectors with this architecture, and the lake-native direction is the most coherent answer in the category to the question of why your vectors should live somewhere separate from your data.

Limitations: Milvus is the most operationally demanding system here to self-host — it expects Kubernetes and a real platform team. Zilliz has raised no funding since a $60 million Series B extension on 24 August 2022, which is a long gap in a capital-intensive category. And Zilliz runs VDBBench, the leaderboard on which Zilliz wins; treat those numbers as a vendor’s own.

Best for: billion-vector workloads, and teams whose vectors should sit alongside an existing Iceberg or Lance lakehouse.

4. Weaviate — best hybrid search

Licence: BSD-3-Clause · Current version: 1.39.x, first released 4 August 2026 · Free tier: permanent, 100,000 objects

Weaviate’s differentiator is that hybrid search — BM25 keyword scoring fused with dense vector similarity, plus metadata filters, in a single query — is a first-class feature available on every plan including the free tier, rather than something you assemble yourself from two systems. For search over documents where exact terms matter (product codes, names, legal citations), that is usually the difference between a working feature and a frustrating one.

The free tier, launched on 17 June 2026, is permanent rather than a time-limited sandbox: one cluster, 100,000 objects, 1 GB memory, 10 GB disk, one collection and up to three tenants, with no credit card. Paid pricing meters three dimensions — vector dimensions, storage and backups. Flex costs $45/month minimum with a 99.5% SLA, billing from $0.00465 per million vector dimensions and $0.12 per GiB stored. Premium starts at $400/month, dropping to $0.003875 per million dimensions on shared clusters and $0.002718 on dedicated, with 99.9% and 99.95% SLAs respectively, plus SSO, PrivateLink, CMEK and HIPAA — note HIPAA is dedicated-tier and AWS-only. Be careful with older write-ups: Weaviate’s October 2025 pricing update announced Flex, Plus and Premium tiers, but the Plus tier no longer appears on the live pricing page, so anything quoting it is stale.

Weaviate’s compression work is the most thorough in the open-source tier: four techniques (product, binary, scalar and rotational quantisation), with 8-bit RQ giving 4x compression at 98–99% recall in Weaviate’s internal testing and 1-bit RQ approaching 32x. Version 1.38, released 5 June 2026, brought HFresh — a disk-based index that keeps only a compressed centroid index in memory — and a built-in MCP server to general availability. Weaviate also shipped Engram, a managed agent-memory service, in June 2026, and its Query Agent costs $30 per organisation per month including 4,000 requests.

Why it wins: the cleanest hybrid search in the category, a permissive BSD-3 licence, and a permanent free tier that is actually usable for a small production workload.

Limitations: dimension-based billing is harder to forecast than per-GB pricing, and it penalises high-dimension embeddings specifically. Funding is opaque — the last verified round is the $50 million Series B of 21 April 2023 led by Index Ventures, and the only confirmed 2026 investment is an undisclosed amount from the RICOH Innovation Fund agreed on 13 March 2026. Circulating “Series C” figures are contradicted by Weaviate’s own materials and should not be relied on.

Best for: document search and RAG where keyword precision matters as much as semantic similarity.

5. turbopuffer — best storage economics at very large scale

Licence: proprietary · General availability: May 2025 · Entry price: $16/month launch tier, no free tier

turbopuffer built vector search directly on object storage with a three-tier cache — cold on S3-class storage, warm on NVMe SSD, hot in memory — and the resulting cost curve is the flattest in the category. It is the system behind an unusually strong customer list: Anthropic, Notion, Cursor, Atlassian, Ramp, Linear, Grammarly and Harvey are all on its public logo wall.

The published case-study numbers are the most striking in this article. Cursor runs over 1 trillion vectors and reports saving 95% on storage and retrieval with over 1 million writes per second. turbopuffer’s Notion case study reports over 10 billion documents across more than 10 million namespaces with 1 GB/s write peaks. Linear reports a 70% cost reduction across 2 billion-plus documents at a 13 ms p50. turbopuffer’s own architecture claims are 1 trillion-plus documents, 10 million-plus writes per second, 25,000-plus queries per second in production, and a maximum of 128 billion documents or 256 TB per namespace.

Pricing has moved sharply in customers’ favour. The launch tier fell from $64 to $16 per month in June 2026; scale is $256/month and adds HIPAA BAA, SSO and audit logs; enterprise starts at $4,096/month with a 35% usage premium and adds single tenancy, BYOC, per-namespace CMEK and a 99.95% SLA. Query pricing was cut by up to 94% in February 2026, from $5 per PB queried to $1 per PB. Per-unit storage and write rates are data not available as static text — turbopuffer publishes them only through a client-side calculator, and the third-party figures in circulation disagree with one another. There is no free tier.

Why it wins: for large, multi-tenant, unevenly-queried corpora, nothing else gets close on cost per vector stored.

Limitations: the cold-path latency is real, and turbopuffer publishes it honestly — on its own 10-million-document, 1,024-dimension benchmark, warm queries return at p50 14 ms / p99 27 ms while cold queries return at p50 874 ms / p99 1,686 ms. If your access pattern does not concentrate, you will feel that. There is also a direct contradiction in the public record worth flagging: turbopuffer’s own benchmark claims 200 ms p99 at over 1,000 QPS with 92% recall across 100 billion vectors, while Zilliz’s VDBBench measures turbopuffer at a 2.34-second max-concurrency P99 and 395.7 QPS on a 100-million-vector dataset. Both are vendor-run, they disagree by an order of magnitude, and no independent test resolves it. Finally, turbopuffer is a 22-person company (as of December 2025) whose December 2025 round from Lachy Groom and Thrive Capital was undisclosed in size.

Best for: search-heavy products with enormous corpora, many tenants and spiky query patterns — and teams comfortable with the concentration risk of a small vendor.

6. Chroma — best for prototyping and local development

Licence: Apache 2.0 · Current version: chromadb 1.5.9, 5 May 2026 · Cloud: generally available since 18 August 2025

Chroma is the shortest path from nothing to working retrieval. pip install chromadb, three lines of Python, and you have a persistent embedded vector store — which is why it has 29,100 GitHub stars and is the default in more RAG tutorials than anything else. The Rust rewrite completed with v1.0.0 on 3 April 2025, which Chroma marketed as a 4x speed-up, so the “toy” reputation from its Python era is now out of date.

Chroma Cloud has been generally available since 18 August 2025 on AWS and GCP, built on a SPANN plus SPFresh distributed index with an object-storage-backed write-ahead log. Its pricing is the most granular in the category: $2.50 per GiB written, $0.33 per GiB stored per month, $0.0075 per TiB queried, $0.09 per GiB egress, and $0.03 per collection fork. Chroma’s own worked example — 13 GiB written, 80 GiB stored, 10 million queries at 4 QPS — comes to $79 per month. Plans are Starter at $0 plus usage with $5 in free credits, Team at $250/month with SOC 2, and Enterprise with BYOC clusters. Customer-managed encryption keys reached general availability on 18 December 2025.

One genuinely distinctive feature: collection forking at $0.03 per request, copy-on-write, which makes branch-per-experiment evaluation workflows cheap in a way no competitor matches.

Why it wins: the lowest friction to a working prototype, an honest Apache 2.0 licence, and a cloud service whose pricing you can actually model line by line.

Limitations: the query billing model is unusual and easy to misread — each metadata or full-text predicate counts as an additional query, and a full-text filter bills as (N − 2) queries where N is the character count of the search string, so a 10-character text filter bills as eight queries rather than one. Chroma is also the least-funded serious vendor here: the only verifiable round is the $18 million seed announced in April 2023, and the several later-round claims in circulation contradict one another and cannot be confirmed.

Best for: prototypes, local development, evaluation harnesses and small production RAG workloads.

7. LanceDB — best for multimodal and lakehouse retrieval

Licence: Apache 2.0 · Current version: lancedb 0.37.1 (10 August 2026); Lance core 10.0.0 (7 August 2026)

LanceDB has repositioned from vector database to “multimodal lakehouse for AI”, and that is the honest description — the Lance columnar format is the product, and vector search is one thing it does. It is the natural fit when your retrieval corpus is images, video and audio alongside text, which is why Midjourney, Runway, Netflix, Uber, NVIDIA, Character.AI and ByteDance appear on its customer list. LanceDB’s own customer page reports over 20,000 vector queries per second at Netflix, and over 90% p90 latency reduction at Character.AI after it migrated full-text search from Elasticsearch to LanceDB.

There is one genuinely independent data point in LanceDB’s favour, which is rare in this category: DuckDB’s own blog on 21 May 2026 measured cold indexed vector queries at 12 ms on Lance against 104 ms for DuckDB’s own index and 761 ms for Parquet. The DuckDB result is a third party publishing a finding that favours a competing format.

Limitations, and they are significant. LanceDB Cloud has effectively disappeared from the official documentation — the docs site’s page index contains no Cloud pages, the FAQ names only the open-source and Enterprise products, and the changelog’s Cloud entries stop in May 2025. Its last verifiable status is public beta; general availability, pricing and free-tier terms are all data not available, and the third-party figures circulating for LanceDB Cloud pricing are uncorroborated. LanceDB Enterprise is the only documented commercial product, and its only public price is an AWS Marketplace annual commit at $60,000 per 12-month contract plus usage. LanceDB’s own comparison table puts the open-source version at 10–50 QPS with 500–1,000 ms object-storage latency against Enterprise at up to 10,000 QPS and 50–200 ms — a large gap that tells you the free version is not the production version. There is also no credible independent ANN benchmark for LanceDB; every “LanceDB QPS at 0.95 recall” figure we could trace led back to AI-generated SEO content with no reproducible harness. LanceDB raised a $30 million Series A on 24 June 2025 led by Theory Ventures, taking total funding to roughly $41 million.

Best for: multimodal corpora and teams already committed to a lakehouse architecture.

8. Vespa — best for complex ranking at scale

Licence: Apache 2.0 · Current platform: 8.738.17, published around 13 August 2026, released Monday to Thursday weekly

Vespa is the most capable and least fashionable system in this article. It is a full search and ranking platform — multi-phase ranking, tensor expressions, machine-learned rankers evaluated per document at query time — that happens to do vector search well. Vespa ran roughly 150 applications serving nearly a billion users at 800,000 queries per second inside Yahoo before spinning out on 4 October 2023, and Perplexity, Spotify, Etsy, AlphaSense and Vinted run it today.

Cloud pricing is straightforward resource billing: Startup at $0.05 per vCPU-hour, Basic at $0.10, Commercial at $0.145 and Enterprise at $0.18, with memory at a tenth of the vCPU rate and unit prices falling linearly with allocated resources up to a 50% reduction. Enterprise carries a $20,000 per month minimum, and enclave BYOC requires a $10,000 per month tenant spend. A free trial gives $300 in credits without a credit card.

Why it wins: if your ranking problem is genuinely complex — business rules, freshness, personalisation and semantic relevance all fused into one scoring function — Vespa is the only system here designed for it rather than retrofitted.

Limitations: the steepest learning curve in the category by a wide margin, and the enterprise minimums put managed Vespa out of reach for most teams. Vespa 9 has not shipped — the release notes still describe it as work in progress, tentatively planned for Q1 2026, and it removes deprecated APIs rather than adding functionality. Funding is a $31 million Series A from Blossom Capital on 1 November 2023 with nothing since.

Best for: large-scale search products where ranking sophistication, not vector storage, is the hard problem.

Vector search in a database you already run

For a large share of teams this section, not the one above, contains the right answer.

pgvector — the default for most teams

Licence: PostgreSQL Licence · Current version: 0.8.2 · Stars: 22,700

pgvector adds vector types and HNSW and IVFFlat indexes to PostgreSQL, and for anything up to roughly ten million vectors it is very hard to argue against. Your vectors sit in transactions alongside the rows they describe, your existing backups cover them, your existing monitoring sees them, and joins and filters are just SQL rather than a metadata query language.

Update urgently if you have not: version 0.8.2, released 25–26 February 2026 (pgvector’s changelog and postgresql.org disagree by a day), fixes CVE-2026-3172, a buffer overflow caused by integer wraparound during parallel HNSW index builds. Any database user able to create or reindex an HNSW index with parallel workers can use it to leak data from other relations or crash the server. Versions 0.6.0 to 0.8.1 are affected. If you cannot patch immediately, setting max_parallel_maintenance_workers to 0 disables parallel HNSW builds and removes the exposure. Upgrading is the single most actionable item on this page.

Know the ceilings before you commit. The vector type stores up to 16,000 dimensions but indexes only 2,000; halfvec raises the indexable limit to 4,000, which is why 3,072-dimension OpenAI embeddings are conventionally cast to halfvec; sparsevec indexes up to 1,000 non-zero elements and only under HNSW. Defaults are m=16, ef_construction=64 and hnsw.ef_search=40. The practical failure mode is index build memory — if the graph exceeds maintenance_work_mem, Postgres logs a notice and the build slows dramatically.

Two extensions push past pgvector’s comfortable ceiling. pgvectorscale from TigerData adds a StreamingDiskANN index and statistical binary quantisation under the PostgreSQL Licence, but note it has had no release since v0.9.0 in November 2025, roughly nine and a half months. VectorChord reached 1.1.1 on 28 February 2026 with RaBitQ compression, dual-licensed AGPLv3 or Elastic License v2, and claims 400,000 vectors per dollar — around 26 times pgvector’s cost efficiency by its own measure, with a 100-million-vector, 768-dimension deployment costed at $247 per month on a single EC2 instance.

The most-cited pgvector performance claim needs a caveat. TigerData’s benchmark reports pgvectorscale beating Pinecone’s storage-optimised s1 tier by 28x on p95 latency and 16x on throughput at 99% recall, at 75% lower cost ($835/month against $3,241) — but it was published in June 2024, and against Pinecone’s performance-optimised p2 tier at 90% recall the same benchmark reports only 1.4x and 1.5x. The headline compares against a competitor’s slowest tier.

Two dead ends to avoid: pgvecto.rs is deprecated, with its own README directing users to VectorChord, and Neon’s pg_embedding was archived in November 2024. Every serious Postgres search extension in 2026 — VectorChord’s BM25 module, ParadeDB’s pg_search (v0.25.3, 17 August 2026, AGPL-3.0), TigerData’s pg_textsearch — still delegates its vector work to pgvector.

Elasticsearch and OpenSearch — best quantisation, and the honest benchmark

Elasticsearch 9.5.0 reached general availability on 4 August 2026, licensed AGPLv3 alongside SSPL and Elastic License 2.0 since 29 August 2024. Its Better Binary Quantization has been the default for dense vectors of 384 dimensions and above since 9.1, delivering roughly 32x compression while, on Elastic’s ten-dataset BEIR testing, beating uncompressed float32 HNSW on NDCG@10 on nine of ten. Elastic states BBQ takes 20–30x less quantisation time than product quantisation and runs queries 2–5x faster. Version 9.5 adds a dedicated VectorDB index mode and DiskBBQ auto-calibration. Elastic Cloud Serverless bills $0.14 per VCU-hour for ingest, $0.09 for search, $0.07 for ML and $0.047 per GB retained per month, with vector profiles receiving 50 GB free; Elastic’s own worked examples put a 2 GB development workload at $24/month and a 20 GB production workload at $190/month.

OpenSearch 3.8.0 shipped on 4 August 2026 under Apache 2.0, governed by the OpenSearch Software Foundation under the Linux Foundation since 16 September 2024, with long-term support announced on 16 April 2026. Its quantisation levels run from 1x to 32x with 32x the default, and OpenSearch reports that the 1-bit scalar quantisation added in 3.6 improved recall by 24% against its previous 32x binary path while cutting latency by 15%. OpenSearch’s own published test on 113 million 1,024-dimension vectors shows recall@100 of 0.95 in memory, 0.98 at 8x compression, 0.97 at 16x and 0.94 at 32x, with p90 latency of 24.02 ms in memory against 96–108 ms on disk — a useful, honestly-reported picture of the disk trade-off. GPU-accelerated indexing via NVIDIA cuVS reached general availability on 2 December 2025 and cut a 10-million-vector, 768-dimension index build from 273.78 minutes to 29.13 minutes at $11.04 down to $2.94. Next-generation Amazon OpenSearch Serverless reached general availability in May 2026 with scale-to-zero after ten minutes idle and up to 60% lower cost; OCU-hours bill at $0.24.

Elastic’s own April 2025 head-to-head measured Elasticsearch 9.0 BBQ at 27.33 ms mean latency and 250.55 operations per second against OpenSearch 2.19’s 156.81 ms and 58.30 — but read the recall column, because OpenSearch consistently landed at 0.96 against Elasticsearch’s 0.95, a real confound in the winner’s favour, and the test is sixteen months old.

Redis — lowest latency when everything fits in RAM

Redis 8.10.0 reached general availability in July 2026, with an 8.10.1 security patch in August. Licensing is the thing to check first: Redis 8.0 and above are tri-licensed under RSALv2, SSPLv1 or AGPLv3, with AGPLv3 added at the Redis 8 launch announced on 1 May 2025 — but 7.4.x through 7.8.x remain dual RSALv2/SSPLv1 only, with no AGPL option. If open-source licensing is a hard requirement, either move to Redis 8 or use Valkey, which reached 9.1 general availability under the Linux Foundation on 19 May 2026 with Valkey Search 1.2 contributed by Google.

The Redis Query Engine supports FLAT, HNSW and SVS-VAMANA vector indexes, and FT.HYBRID — reciprocal rank fusion and linear combination of keyword and vector scores — shipped in Redis 8.4. Redis’s own test indexed 1 billion 768-dimension FLOAT16 vectors with sub-second median latency under concurrent query load. Redis Cloud is free to 30 MB, with Essentials from $0.007 per hour (about $5/month) and Pro from $0.014 per hour with a $200/month minimum.

Redis’s competitive benchmark — up to 3.4x the queries per second of Qdrant, 3.3x Milvus and 1.7x Weaviate at recall ≥ 0.98 — is from June 2024, run on Redis 7.4, and used a scaling factor of 4 on nominally matched hardware. Redis also discloses that its MongoDB comparison only returned results at a lower recall level on one dataset. Redis’s competitive benchmark is more than two years stale and should not decide anything today.

MongoDB Vector Search — if you are already on Atlas

MongoDB Vector Search has been generally available since 4 December 2023 and became available for self-managed Enterprise Advanced and Community deployments on 1 July 2026, requiring MongoDB 8.2 or later, with the mongot process source-available under SSPL. Vector queries are included in cluster cost, but dedicated Search Nodes bill separately by the hour — from $0.12 for an S20 on AWS to $3.26 for an S80, with a minimum of two and a maximum of 32. The free M0 tier gives 512 MB and a maximum of three indexes; Atlas Flex runs from about $8 to a hard-capped $30 per month.

Maximum dimensions are 8,192 (raised from 4,096 on 30 March 2025) with a ceiling of 2.1 billion index objects per index. Automatic quantisation reached general availability on 29 April 2025 and cuts search-node RAM by 3.75x for scalar and 24x for binary. MongoDB’s own benchmark on 15.3 million vectors using voyage-3-large at 2,048 dimensions with quantisation enabled reports 90–95% accuracy at under 50 ms query latency. The Voyage AI acquisition — announced 24 February 2025, reported by Bloomberg at roughly $220 million though never confirmed by MongoDB — has since produced Voyage 4 embeddings in January 2026, native reranking via $rerank in June 2026, and voyage-code-4 in August 2026, making Atlas the most vertically-integrated retrieval stack of the incumbents.

Amazon S3 Vectors — cheapest cold storage

Amazon S3 Vectors went to preview on 15 July 2025 and generally available on 2 December 2025, expanding from 14 to 34 Regions by August 2026. Pricing is $0.06 per GB-month of logical vector storage, $0.20 per GB uploaded (with a 128 KB minimum per PUT), $2.50 per million query requests, plus data-processed charges from $0.004 per TB down to $0.0004 per TB above 10 million vectors, and $0.01 per GB returned. AWS cut data-processed query charges on indexes over 10 million vectors by up to 80% on 16 June 2026 — from $0.002 to $0.0004 per TB — and the reduction applied automatically. AWS claims up to 90% lower total cost than specialised vector databases.

The constraints are real and mostly immutable once set: maximum 2 billion vectors per index, 10,000 indexes per bucket, 1 to 4,096 dimensions, Cosine or Euclidean distance only, 40 KB of metadata per vector of which 2 KB is filterable, and 1,000 write requests per second per index. AWS’s own latency guidance is the honest part: “infrequent queries continue to return results in under one second, with more frequent queries now resulting in latencies around 100 milliseconds or less.” That is not an interactive-search latency budget. S3 Vectors is best understood as the cold tier of a retrieval architecture — it integrates natively with Bedrock Knowledge Bases and supports tiering from Amazon OpenSearch Service.

Pricing compared: what you will actually pay

Entry price is the lowest recurring commitment for a production-grade deployment. Because these systems meter completely different things — dimensions, read units, compute units, GB-hours, PB queried — a single per-vector figure would be misleading, so the metering basis is shown instead.

The cost lever that matters most is not the vendor — it is quantisation. Moving from float32 to 8-bit scalar quantisation cuts vector memory by roughly 4x at recall costs most applications cannot detect; binary quantisation cuts it by 24 to 32 times. MongoDB measures a 24x search-node RAM reduction from binary quantisation; Weaviate measures 98–99% recall retained at 4x compression with 8-bit rotational quantisation. Every system in this table supports some form of it. Configure it before you shop for a cheaper vendor.

The second lever is access pattern. If your queries are infrequent or concentrate on a small hot subset, object-storage-native systems (turbopuffer, S3 Vectors, Zilliz tiered storage) cost a fraction of memory-resident ones. If every query must return in under 50 milliseconds regardless of what was queried last, you are paying for RAM and no architecture avoids that.

Benchmarks: what is actually trustworthy

This is the least reliable part of the vector database category, and it is worth being blunt about why.

ANN-Benchmarks, the long-standing reference, is no longer maintained. Its README now opens by saying so and directing readers elsewhere; its last results run was in April 2025.

VIBE is the credible independent replacement. Published as arXiv:2505.17810 on 23 May 2025 by researchers at Helsinki, Aalto, Padova and the IT University of Copenhagen — including Martin Aumüller, a co-author of the original ANN-Benchmarks paper — it covers 21 implementations across 12 in-distribution and 6 out-of-distribution datasets. Its headline findings: at 95% recall, SymphonyQG has the highest throughput on 5 of 12 datasets and Glass on 4 of 12; graph and clustering methods significantly outperform hashing and tree methods; binarisation delivers roughly 4x throughput (about 2,000 QPS float against 8,000 QPS binarised on one dataset at 90% recall); CAGRA is fastest on GPU; and out-of-distribution queries remain the open problem, with the fastest in-distribution methods degrading significantly on them. The important caveat: VIBE benchmarks indexing algorithms, not databases. It tells you nothing about filtering, durability, multi-tenancy or operational cost.

The only cross-vendor 2026 numbers come from VDBBench, which Zilliz runs. Zilliz wins its own leaderboard. Read it with that in mind, and note that benchANT maintains an independent fork of the harness, which is the closest thing to a neutral operator in this space. On a 1-million-vector Cohere dataset at a $1,000/month cost budget, VDBBench reports: Zilliz Cloud 8cu-perf at 13,316 QPS / 0.9383 recall / 2 ms P99; Milvus sq8 at 5,973 QPS / 0.9192; OpenSearch at 3,055 / 0.9066; Elastic Cloud at 2,354 / 0.9143; Qdrant Cloud at 1,242 / 0.9474; Pinecone p2.x8 at 1,147 / 0.9262. On its May 2026 cost leaderboard over 100 million LAION vectors: Zilliz Capacity 32CU at 786 QPS with a 158 ms P99 and 0.9728 recall; turbopuffer at 396 QPS with a 2.34-second P99; Pinecone Serverless at 4.6 QPS with a 4.85-second P99 at 0.9609 recall, with ingestion write costs of $9.12, $302 and $1,180 respectively.

The most instructive benchmark of 2026 was the Elastic–Qdrant exchange, because it ended in a retraction. All three rounds used the same 21,015,300-passage Wikipedia DPR dataset with 768-dimension e5-base-v2 embeddings, measuring recall@100 against prepared ground truth. Elastic published first, measuring Elasticsearch bbq_disk at 32.4 QPS against Qdrant’s 4.5 QPS at recall around 0.96. Qdrant rebutted on 8 July 2026 with a two-stage prefetch-and-rescore configuration using io_uring, reporting 111.9 QPS at 35.7 ms on smaller hardware at the same 0.9596 recall. Elastic then reproduced Qdrant’s configuration on 13 July 2026 and published findings against its own original claim:

• The io_uring async scorer moved almost nothing — 31.6 QPS to 35.8 QPS, which Elastic described as “within run-to-run noise”. Docker’s default seccomp profile had been silently blocking io_uring entirely.
• The “disk-based” benchmark never touched the disk. During the measurement window iostat showed 0 MB/s read and 0 IOPS across all three nodes while CPU ran at 60–100%. The working set was in page cache.
• Segment count, not tuning, explained the gap: 128 segments gave 35.8 QPS at 0.9745 recall; 67 segments gave 67.2 QPS at 0.9596. Roughly 2x throughput came purely from segment count, with recall falling as throughput improved.
• The memory dispute was an accounting artefact — Qdrant had compared Elasticsearch’s JVM heap against Qdrant’s total process memory.
• Elastic retracted its own original explanation: “the mechanism the original post named, random disk reads during rescore, does not survive the same arithmetic.”

Elastic’s conclusion was that at around 0.96 recall on identical hardware, the two systems land at roughly 56 QPS and 53–67 QPS respectively — effectively a tie, after two vendors had published numbers implying gaps of roughly 7x and 3.5x in opposite directions.

How to read a vector database benchmark

1. Check the recall column first. A system 5x faster at 0.94 recall than another at 0.98 has not won anything. Two results are comparable only at similar precision.
2. Check the date. Qdrant’s public benchmark is frozen at June 2024, Redis’s at June 2024, TigerData’s at June 2024, ANN-Benchmarks at April 2025. Only VDBBench and the 2026 Elastic–Qdrant exchange carry current numbers.
3. Check who ran it. Every headline figure in this category except VIBE and the DuckDB Lance test is vendor-run.
4. Check the tier being compared against. TigerData’s famous 28x is against Pinecone’s storage-optimised tier; against the performance tier the same benchmark reports 1.4x.
5. Check segment and shard counts. They are the dominant hidden variable and are rarely reported.
6. Distrust precise-looking tables with no harness. A large share of “vector database benchmark 2026” content publishes reproducible-looking numbers that trace to no primary source at all — we could not find a single credible independent benchmark for LanceDB, for instance, despite dozens of pages claiming specific QPS figures.

Best for each situation

For most teams starting today

Winner: pgvector on the PostgreSQL you already run

Under roughly ten million vectors, the operational savings of one fewer system outweigh any performance advantage a specialist offers. Your vectors stay transactional, your backups already cover them, and filters are just SQL. Upgrade to 0.8.2 for the CVE-2026-3172 fix. Move on when index builds start dominating your maintenance window or filtered queries stop meeting latency targets — then try pgvectorscale or VectorChord before you migrate.

For the best dedicated vector database

Winner: Qdrant

Apache 2.0, the easiest dedicated engine to self-host, the strongest filtering in the open-source tier, current quantisation via TurboQuant in 1.19.0, a permanent free managed tier, and a $50 million Series B in March 2026 behind it. Alternative: Weaviate if hybrid keyword search is central to your product.

For fully managed with no platform team

Winner: Pinecone

Nothing to size, nothing to tune, and the Builder plan at $20/month blocks rather than bills overage — the cleanest spend cap in the category. Caveat: model your cost at target scale first, and factor in the unresolved corporate situation.

For billions of vectors

Winner: Milvus self-hosted, or Zilliz Cloud managed

The only open-source architecture in the category genuinely designed for this, and Milvus 3.0’s external collections index Iceberg, Lance and Parquet in place rather than duplicating them. Alternative: turbopuffer if the corpus is enormous but query volume is uneven.

For hybrid keyword and vector search

Winner: Weaviate

BM25 and dense vectors fused in a single query, on every plan including the permanent free tier. Alternatives: Elasticsearch or Redis 8.4’s FT.HYBRID if you already run either.

For the lowest cost per vector stored

Winner: Amazon S3 Vectors at $0.06 per GB-month

No floor, no cluster, 34 Regions, and native Bedrock Knowledge Base integration. Accept the latency: AWS’s own guidance is around 100 ms for frequent queries and under a second for infrequent ones. Alternative: turbopuffer from $16/month when you need real interactive latency on the warm path.

For very large multi-tenant search products

Winner: turbopuffer

Cursor, Notion, Linear and Grammarly all run on it, with published savings of 70–95% against their previous stacks. Caveat: cold-path p50 is 874 ms on turbopuffer’s own benchmark, and it is a 22-person company.

For prototyping and local development

Winner: Chroma

Fastest path from nothing to working retrieval, Apache 2.0, Rust core since v1.0.0, and collection forking at $0.03 makes branch-per-experiment evaluation cheap. Alternative: LanceDB if your data is multimodal.

For multimodal retrieval

Winner: LanceDB

The Lance format handles images, video and audio alongside text natively, with an independent DuckDB test measuring 12 ms cold indexed queries against 761 ms for Parquet. Caveat: LanceDB Cloud’s status is unclear and the open-source tier is rated at 10–50 QPS by LanceDB itself.

For complex ranking at scale

Winner: Vespa

Multi-phase ranking with machine-learned rankers evaluated per document, proven at 800,000 queries per second inside Yahoo. Caveat: the steepest learning curve here, and Enterprise starts at $20,000 per month.

For the strictest licence requirements

Winner: Qdrant, Milvus, Chroma, LanceDB or OpenSearch — all Apache 2.0

Weaviate is BSD-3-Clause, which is equally permissive. Watch out for Redis 7.4–7.8 (RSALv2/SSPLv1 only, no open-source option), VectorChord and ParadeDB (AGPL), and MongoDB’s mongot (SSPL).

How to choose a vector database

Seven questions decide this, and they should be answered in this order.

1. How many vectors, honestly, in twelve months? Under a million, use whatever is already in your stack. One to ten million, pgvector is comfortable. Ten to a hundred million, a dedicated database starts earning its keep. Above that, you are choosing between Milvus, turbopuffer and a serious managed contract.

2. Do you need filtering, and how selective? Filtered vector search is where implementations diverge most. Qdrant, Weaviate and Milvus handle high-selectivity filters well; naive implementations either scan everything or return too few results. Test this with your real filter distribution, not a synthetic one.

3. Do you need keyword search too? If exact terms matter — product codes, names, citations — you need hybrid search, and assembling it yourself from two systems is more work than it appears. Weaviate, Elasticsearch, OpenSearch, Vespa and Redis 8.4 all do it natively.

4. What is your latency budget, and for which queries? A uniform sub-50 ms requirement means paying for memory. If 95% of queries hit a hot subset, tiered and object-storage-native architectures cut cost dramatically for the same felt experience.

5. Who operates it? Self-hosting Qdrant or Chroma is a modest commitment. Self-hosting Milvus or Vespa is a platform-team commitment. If you have neither, the managed floor on the dedicated vector databases runs from $16 to $500 per month before storage, with no floor at all on purely usage-billed services and five figures a month on Vespa and LanceDB enterprise contracts.

6. What is your licence and compliance floor? Apache 2.0 and BSD-3 systems carry no relicensing risk. If you need HIPAA, check the tier — it is a $190/month add-on on Pinecone Standard, dedicated-and-AWS-only on Weaviate, and included on Pinecone Enterprise and Qdrant.

7. What happens if the vendor disappears? This is not hypothetical here. MyScale shut down in May 2025, SingleStore was taken private in October 2025, Marqo pivoted in May 2026, and Vald has been dormant since July 2025. An Apache-2.0 system you could run yourself is meaningfully cheaper insurance than a migration plan.

A closing note on scope: this page ranks the systems that store and search vectors. Which model produces those vectors is a separate decision with a bigger effect on retrieval quality than the database choice — quantisation aside, a better embedding model will improve your results more than a faster index will. Our picks here are weighted for operational fit and verifiable cost, and cross-checked against the same evidence behind our best LLM APIs ranking.

Frequently asked questions

What is a vector database?

A vector database stores high-dimensional numerical arrays called embeddings and finds the ones most similar to a query embedding, usually using approximate nearest-neighbour indexes such as HNSW, IVF or DiskANN. It is the retrieval layer behind semantic search, recommendation systems and retrieval-augmented generation, where the goal is finding text, images or audio that mean something similar rather than matching keywords exactly. In 2026 the term covers two different things: purpose-built systems like Qdrant, Milvus and Pinecone, and vector indexes added to general-purpose databases like PostgreSQL, Redis, MongoDB and Elasticsearch.

What is the best vector database in 2026?

Qdrant is the best all-round dedicated vector database — Apache 2.0, the easiest of the dedicated engines to self-host, strong metadata filtering, a permanent free managed tier, and a $50 million Series B closed in March 2026. Pinecone is the best fully managed option at a $50/month Standard minimum. Milvus is the best at billion-vector scale. Weaviate is the best for hybrid keyword-plus-vector search. There is no single winner, because the right answer depends on scale, whether you can operate it and whether retrieval is your product or a feature of it.

Do I actually need a vector database?

Probably not, if you are under roughly ten million vectors and already run PostgreSQL. pgvector adds vector types and HNSW indexes to Postgres, keeps your embeddings transactional and inside your existing backups, and lets you filter with ordinary SQL — and every serious Postgres search extension in 2026 still delegates its vector work to it. The case for a dedicated system strengthens when index builds start dominating your maintenance window, when filtered queries stop meeting latency targets, or when you pass roughly 100 million vectors. Redis, MongoDB, Elasticsearch and OpenSearch users have the same option inside the system they already run.

What is the cheapest vector database?

Amazon S3 Vectors is the cheapest per vector stored at $0.06 per GB-month with no minimum spend, though queries return in roughly 100 milliseconds at best and under a second when infrequent, by AWS’s own guidance. If you already run PostgreSQL, pgvector is cheaper still because it costs nothing beyond the database you already pay for. Among managed dedicated services, turbopuffer has the lowest entry price at $16 per month (cut from $64 in June 2026), followed by Pinecone Builder at $20/month and Weaviate Flex at $45/month. Chroma Cloud has no monthly floor at all, billing $2.50 per GiB written and $0.33 per GiB stored.

Which vector databases are open source?

Qdrant, Milvus, Chroma, LanceDB, Vespa and OpenSearch are all Apache 2.0. Weaviate is BSD-3-Clause, equally permissive. pgvector uses the PostgreSQL Licence. Watch the qualified cases: Elasticsearch is tri-licensed AGPLv3, SSPL and Elastic License 2.0; Redis 8.0 and above are tri-licensed RSALv2, SSPLv1 and AGPLv3, but Redis 7.4 to 7.8 are RSALv2/SSPLv1 only with no open-source option; VectorChord and ParadeDB are AGPL; and MongoDB’s mongot search process is source-available under SSPL rather than open source. Pinecone and turbopuffer are proprietary with no self-hosted open-source version.

Is pgvector good enough for production?

Yes, for most production workloads. pgvector 0.8.2 supports HNSW and IVFFlat indexes, iterative scans that make filtered queries return complete results, and half-precision and binary vector types. Its practical limits are the ones to plan around: it indexes 2,000 dimensions on the vector type and 4,000 on halfvec (which is why 3,072-dimension OpenAI embeddings are cast to halfvec), and index build performance degrades sharply once the graph exceeds maintenance_work_mem. Upgrade to 0.8.2 if you have not — it fixes CVE-2026-3172, a buffer overflow during parallel HNSW index builds that can leak data from other relations or crash the server. Beyond pgvector’s comfortable ceiling, pgvectorscale adds a StreamingDiskANN index and VectorChord adds RaBitQ compression, both without leaving Postgres.

How much does Pinecone cost?

Pinecone has four plans as of August 2026: Starter is free with 2 GB storage, 1 million read units and 2 million write units per month; Builder is a flat $20/month where usage beyond plan limits is blocked rather than billed; Standard carries a $50/month minimum usage commitment; and Enterprise carries $500/month with a 99.95% SLA, BYOC and CMEK. Serverless usage bills at $0.33 per GB per month for storage and $16 per million read units in AWS us-east-1; the write-unit rate is not published as static text on Pinecone’s site. A query costs 1 read unit per GB of namespace size with a 0.25 RU minimum, so namespace size — not top_k — drives query cost. HIPAA is a $190/month add-on on Standard and included on Enterprise.

What is the difference between a vector database and vector search in Postgres?

A dedicated vector database is built around the approximate nearest-neighbour index, so it typically offers better filtered-search performance, distributed scaling past hundreds of millions of vectors, and features like multi-tenancy, hybrid ranking and quantisation tuned specifically for vectors. Vector search in PostgreSQL via pgvector puts the index inside a general-purpose transactional database, so embeddings stay consistent with the rows they describe, existing backups and monitoring cover them, and filters are ordinary SQL joins rather than a separate metadata language. The trade is operational simplicity against peak performance: below roughly ten million vectors the simplicity usually wins, and above roughly a hundred million the performance usually does.

Which vector database is fastest?

There is no trustworthy answer, and the honest reason is worth knowing. Almost every published figure in this category is vendor-run, and the one 2026 benchmark that was carefully reproduced collapsed under it — Elastic’s 13 July 2026 reproduction of its own Elasticsearch-versus-Qdrant comparison found that the disk-based test never read from disk (iostat showed 0 IOPS throughout), that the io_uring feature credited with the difference was silently blocked by Docker’s seccomp profile, and that segment count alone explained roughly 2x of the throughput gap. Elastic’s revised conclusion was that both systems land at roughly 53–67 queries per second at 0.96 recall on identical hardware. For algorithm-level comparisons, the independent VIBE benchmark is the best current source, though it measures indexing algorithms rather than databases.

What is hybrid search, and which vector databases support it?

Hybrid search combines keyword scoring (usually BM25) with dense vector similarity in a single query, fusing the two rankings — most often with reciprocal rank fusion — so exact terms like product codes, names and citations are not lost to semantic approximation. Weaviate offers it on every plan including its free tier and is the strongest implementation among dedicated vector databases. Elasticsearch, OpenSearch and Vespa have mature hybrid ranking as search engines first. Redis added FT.HYBRID with reciprocal rank fusion and linear combination in version 8.4. Pinecone supports dense-plus-sparse in one index when the metric is dotproduct, and now offers full-text as a first-class index type.

How much can quantisation reduce vector database costs?

Substantially, and it is the highest-leverage change available. Scalar quantisation to 8-bit cuts vector memory by roughly 4x; binary quantisation to 1 bit per dimension cuts it by 24 to 32 times. MongoDB measures search-node RAM falling 3.75x with scalar and 24x with binary quantisation. Weaviate reports 98–99% recall retained at 4x compression with 8-bit rotational quantisation. In Elastic’s own ten-dataset BEIR testing, Better Binary Quantization beat uncompressed float32 HNSW on NDCG@10 on nine of ten datasets while delivering roughly 32x compression. OpenSearch’s own testing on 113 million vectors shows recall@100 of 0.95 uncompressed against 0.94 at 32x compression. Configure quantisation before you shop for a cheaper vendor — it usually saves more than switching does.

Has any vector database company shut down or been acquired?

Yes. MyScale ceased operations on 28 May 2025. SingleStore was taken private by Vector Capital in a deal that closed on 16 October 2025, reported at roughly $500 million though never officially disclosed. Marqo pivoted away from general vector search to AI-native e-commerce search, launching its Sibbi agent on 7 May 2026. Vald has been dormant since its last release on 4 July 2025. Pinecone remains independent — a sale process was reported by Calcalist on 31 August 2025 but no acquisition has occurred, and Ash Ashutosh replaced founder Edo Liberty as CEO on 8 September 2025. Notably, no vector database was acquired by a hyperscaler in 2025 or 2026; the database acquisitions of that period went to Postgres companies instead, with Databricks buying Neon and Snowflake buying Crunchy Data.

Related pages

Build with AI: Best LLM APIs · Best LLM observability tools · Best AI agent frameworks · Best AI for coding

Run it yourself: Best open-source LLMs · Best local LLMs

Understand the field: Best AI models · What is agentic AI · Best AI for data analysis

Prices, licences, versions and limits were read from each vendor’s own documentation and pricing pages on 19 August 2026. Where a vendor publishes rates only through a client-side calculator, we have written “data not available” rather than repeat unverified third-party figures — this applies to turbopuffer’s per-unit storage and write rates, Pinecone’s write-unit rate and LanceDB Cloud’s pricing. Every performance figure is attributed to whoever ran it, and vendor-run benchmarks are labelled as such. Rate cards in this category change frequently; confirm anything you are budgeting against.