SSubrupt
Skip to content

Best Vector Database for RAGs of 2026

Updated · 4 picks · live pricing · affiliate disclosure

Postgres extension adding vector search to Postgres on Supabase, Neon, RDS.

BEST OVERALL7.0/10Save $36/yr

pgvector (Postgres)

Postgres extension adding vector search to Postgres on Supabase, Neon, RDS.

OSS Apache 2.0 free; Supabase free permanent
Try pgvector (Postgres)See full review

How it stacks up

  • OSS extension free

    vs Pinecone serverless

  • Supabase free 500MB

    vs Qdrant Rust OSS

  • Neon $19/mo upgrade

    vs Weaviate hybrid

#2
Weaviate6.7/10

From $25/mo

View
#3
Qdrant6.4/10

From $25/mo

View

All picks at a glance

#PickBest forStartingScore
1pgvector (Postgres)Best RAG for Postgres teams, vector search beside relational data$19.00/mo7.0/10
2WeaviateBest hybrid search RAG, vector plus BM25 keyword for retrieval quality$25.00/mo6.7/10
3QdrantBest OSS RAG pick, filter-aware HNSW for metadata filtering$25.00/mo6.4/10
4PineconeBest mainstream RAG default, serverless pay-as-you-go for production$500.00/mo5.1/10

Quick pick by use case

If you only have thirty seconds, find your situation below and skip to that pick.

If I want fully-managed serverless RAG with mainstream LangChain integrationPineconePinecone Starter free; Standard pay-as-you-go; Enterprise $500/mo. Mainstream RAG default.If I need OSS self-host with metadata filtering for multi-tenant RAGQdrantQdrant OSS Apache 2.0 free; Free Cluster 1GB; Standard $25/mo. Filter-aware HNSW.If I am running RAG against existing Postgres relational datapgvector (Postgres)pgvector OSS extension free; Supabase free 500MB; Neon $19/mo. Vector beside relational.If I need hybrid vector and BM25 keyword retrieval for quality liftWeaviateWeaviate OSS BSD-3 free with hybrid search; Sandbox 14d cloud; Standard $25/mo.

Compare all 4 picks

Top spec
#1pgvector (Postgres)7.0/10$19.00/mo$228.00/yrSave $36/yrOSS extension free
#2Weaviate6.7/10$25.00/mo$300.00/yr$36/yr moreOSS BSD-3 free
#3Qdrant6.4/10$25.00/mo$300.00/yr$36/yr moreOSS Apache 2.0 free
#4Pinecone5.1/10$500.00/mo$6,000.00/yr$5,736/yr moreStarter free
#1

pgvector (Postgres)

7.0/10Save $36/yr

Best RAG for Postgres teams, vector search beside relational data

Try pgvector (Postgres)See pgvector (Postgres) alternatives

Postgres extension adding vector search to Postgres on Supabase, Neon, RDS.

PlanMonthlyAnnualWhat you get
OSS extensionFree—Postgres extension with HNSW and IVFFlat indexes for hybrid SQL plus vector queries.
Supabase FreeFree—500MB Postgres with pgvector for up to ~5M small vectors free permanently.
Neon Launch$19.00/mo$228.00/yr$19 monthly with 10GB autoscaling Postgres and pgvector built in.
AWS RDS / AuroraFree$0.00/yrStandard RDS pricing with pgvector preinstalled for combined relational data.

pgvector is the Postgres-bundled RAG pick and the right call for teams running RAG against existing relational data. Open-source under Apache 2.0 since 2021. The wedge for RAG readers: vector search adds to the Postgres database the team already runs, the only catalog RAG pick where vector retrieval combines with relational data joins in one SQL query for relational-RAG patterns that pure-vector DBs require application-level merging to achieve.

OSS extension ships free as a Postgres extension on any deployment. Supabase Free covers five hundred megabyte Postgres with pgvector for around five million small vectors. Neon Launch is the upgrade tier at nineteen dollars monthly with ten gigabyte autoscaling Postgres and pgvector built in. AWS RDS ships pgvector preinstalled. Most relational-data RAG teams stay on Supabase Free or self-hosted Postgres until vector volume crosses around fifty million vectors.

The trade-off versus Pinecone is feature breadth; pgvector lacks managed serverless and dedicated vector-DB observability. The trade-off versus Qdrant is performance ceiling at scale; pgvector handles up to around fifty million vectors comfortably. For teams running RAG against existing Postgres relational data, pgvector is the right call.

Pros

  • Vector search lives alongside relational data in one Postgres database
  • Hybrid SQL plus vector queries combine in one query for relational-RAG patterns
  • Free on Supabase 500MB tier with around five million small vectors
  • Neon Launch upgrade at nineteen dollars monthly with autoscaling Postgres
  • Apache 2.0 OSS with broad ecosystem and mature production deployments

Cons

  • Scale ceiling around fifty million vectors; dedicated DBs win at higher scale
  • No native managed-vector-DB features beyond what Postgres provides
OSS extension freeSupabase free 500MBNeon $19/mo upgradeOSS Apache 2.0 free; Supabase free permanent

Best for: RAG teams running retrieval against existing Postgres relational data who want vector search in the same database.

OSS license & sovereignty
9
Query performance
8
Setup complexity
9
Value
10
Support
8
Try pgvector (Postgres)
#2

Weaviate

6.7/10$36/yr more

Best hybrid search RAG, vector plus BM25 keyword for retrieval quality

Try WeaviateSee Weaviate alternatives

Hybrid search OSS with GraphQL API and BSD-3 self-hosting plus managed cloud.

PlanMonthlyAnnualWhat you get
OSS self-hostedFree—BSD-3 licensed self-hosting with multi-vector and hybrid search.
Sandbox (Cloud)Free—Free 14-day trial of single sandbox cluster with all features unlocked.
Standard (Cloud)$25.00/mo$300.00/yr$25 monthly entry with pay-per-dimension and built-in hybrid search.
EnterpriseCustomCustomCustom pricing with BYOC option, multi-region, and premium support.

Weaviate is the hybrid search RAG pick and the right call for production teams that need vector plus keyword retrieval combined for quality lift on mixed text-and-metadata data. Founded 2019 in Amsterdam. The wedge for RAG readers: hybrid (vector plus BM25 keyword) search ships built-in under BSD-3 self-host with GraphQL API, an architectural feature that delivers retrieval quality lift on legal docs, technical documentation, and structured chunks where keyword precision matters as much as semantic similarity.

OSS self-hosted ships free under BSD-3. Sandbox cloud ships fourteen-day trial with all features. Standard cloud is the upgrade tier at twenty-five dollars monthly with hybrid search built in. Enterprise ships custom pricing with BYOC option plus multi-region. Most production RAG teams that valued retrieval quality land on Standard cloud or OSS self-host based on SRE capacity.

The trade-off versus Pinecone is mainstream brand recognition; Weaviate has smaller mainstream brand than Pinecone. The trade-off versus Qdrant is OSS license posture; Weaviate is BSD-3 where Qdrant is Apache 2.0. For RAG teams that need hybrid retrieval quality, Weaviate is the right call.

Pros

  • Hybrid vector and BM25 keyword search built in for retrieval quality lift
  • GraphQL API plus REST for JavaScript-first teams and modern app integration
  • BSD-3 OSS self-host with no licensing cost for unlimited deployment
  • Standard cloud at twenty-five dollars monthly is the cheapest hybrid-search managed entry
  • Founded 2019 in Amsterdam with strong enterprise traction

Cons

  • BSD-3 license posture differs from Apache 2.0 picks for some procurement teams
  • Pay-per-dimension cloud pricing complex for budgeting at scale
OSS BSD-3 freeSandbox 14d trialStandard $25/mo upgradeOSS BSD-3 free; 14-day cloud sandbox

Best for: RAG teams that need hybrid vector and keyword retrieval for quality lift on mixed text-and-metadata data.

OSS license & sovereignty
9
Query performance
8
Setup complexity
8
Value
9
Support
8
Try Weaviate
#3

Qdrant

6.4/10$36/yr more

Best OSS RAG pick, filter-aware HNSW for metadata filtering

Try QdrantSee Qdrant alternatives

Rust-based vector DB with fastest filter-aware HNSW indexing under Apache 2.0.

PlanMonthlyAnnualWhat you get
OSS self-hostedFree—Apache 2.0 Rust-based vector DB with filter-aware HNSW indexing.
Free Cluster (Cloud)Free—1GB always-free production-ready cluster with low replication.
Standard (Cloud)$25.00/mo$300.00/yr$25 monthly for 4GB at $0.108 per GB-hour with higher availability.
Enterprise / HybridCustomCustomBYOC option with multi-region, SSO, and premium support.

Qdrant is the OSS RAG pick and the right call for teams that want metadata filtering during retrieval without paying for managed cloud. Founded 2021 in Berlin. The wedge for RAG readers: filter-aware HNSW indexing maintains sub-millisecond latency when queries combine vector similarity with metadata filters, the only catalog RAG pick where multi-tenant RAG, time-bound retrieval, or source-filtered queries do not degrade query latency under load.

OSS self-hosted ships free under Apache 2.0 with full Rust engine. Free Cluster cloud ships always-free production-ready with one gigabyte storage. Standard cloud is the upgrade tier at twenty-five dollars monthly for four gigabytes. Enterprise ships custom pricing with BYOC option. Most production RAG teams land on Free Cluster for evaluation then pick OSS self-host or Standard cloud based on SRE capacity.

The trade-off versus Pinecone is mainstream brand recognition; Qdrant has smaller ecosystem than Pinecone for tooling integrations. The trade-off versus Weaviate is hybrid search; Qdrant ships filter-aware HNSW where Weaviate ships vector plus BM25 hybrid. For OSS RAG teams with metadata filtering, Qdrant is the right call.

Pros

  • Filter-aware HNSW maintains sub-millisecond latency under metadata-filtered queries
  • Apache 2.0 OSS self-host with no licensing cost for unlimited deployment
  • Free Cluster cloud always-free production-ready with one gigabyte storage
  • Standard cloud at twenty-five dollars monthly with four gigabytes is competitive
  • Founded 2021 in Berlin with strong production-ready positioning

Cons

  • Smaller ecosystem than Pinecone for tooling integrations
  • No GraphQL API; REST and gRPC only for application integration
OSS Apache 2.0 freeFree Cluster 1GBStandard $25/mo upgradeOSS Apache 2.0 free; Free Cluster always-free

Best for: OSS RAG teams that need metadata filtering during retrieval without paying for managed cloud at production scale.

OSS license & sovereignty
9
Query performance
10
Setup complexity
8
Value
9
Support
7
Try Qdrant
#4

Pinecone

5.1/10$5,736/yr more

Best mainstream RAG default, serverless pay-as-you-go for production

Try PineconeSee Pinecone alternatives

Largest serverless vector database for production AI workloads with pay-as-you-go pricing.

PlanMonthlyAnnualWhat you get
Starter (free)Free—5 serverless indexes with 2M reads and 1M writes monthly for evaluation.
Standard (Serverless)Free$0.33/yrPay-as-you-go at ~$0.33 per 1M reads and $0.33 per GB storage.
Enterprise$500.00/mo$6,000.00/yr$500 monthly minimum with volume discounts, SSO, and premium support.
Dedicated (Pod-based legacy)Free$0.00/yrReserved capacity with predictable costs and multi-region replication.

Pinecone is the mainstream RAG default and the right call for production teams that want fully-managed serverless without infrastructure overhead. Founded 2019 by ex-AWS engineer Edo Liberty and backed by a16z. The wedge for RAG readers: serverless pay-as-you-go pricing scales from zero idle with the largest mainstream brand recognition for production AI, and integrations with LangChain, LlamaIndex, CrewAI ship Pinecone as the default during initial RAG selection.

Starter ships free with five serverless indexes plus two million reads and one million writes monthly. Standard Serverless ships pay-as-you-go at fractional cents per million reads and per million writes plus storage. Enterprise ships at five hundred dollars monthly minimum with volume discounts plus SAML SSO. Dedicated Pod-based legacy ships reserved capacity. Most production RAG teams land on Standard Serverless until query volumes cross around one hundred million queries monthly.

The trade-off versus Qdrant is OSS posture; Pinecone is closed-source SaaS where Qdrant ships Apache 2.0 self-host. The trade-off versus pgvector is database integration; Pinecone is standalone where pgvector lives in Postgres. For mainstream production RAG teams, Pinecone is the right call.

Pros

  • Largest mainstream brand for production RAG with the strongest LangChain integration
  • Starter free with five indexes plus generous monthly read and write limits
  • Pay-as-you-go pricing scales from zero idle with no infrastructure overhead
  • Multi-region replication on Dedicated tier for production reliability
  • Founded 2019 with a16z backing and broad ecosystem traction

Cons

  • Cost surprise risk at scale past around one hundred million queries monthly
  • No OSS self-hosting option; managed cloud only
Starter freeStandard PAYGEnterprise $500/mo+Starter free permanent; cancel-anytime

Best for: Production RAG teams that want fully-managed serverless with mainstream brand recognition across LangChain, LlamaIndex, CrewAI integrations.

OSS license & sovereignty
8
Query performance
9
Setup complexity
10
Value
7
Support
8
Try Pinecone

How we picked

Each pick gets a transparent composite score from price, features, free-tier availability, and editor fit. Pricing flows from our live database, so when a vendor changes prices the score updates here too.

We weight price at 40 percent, features at 30, free tier at 15, fit at 15. Pinecone leads because the mainstream serverless model matches production RAG default selection with brand recognition across LangChain, LlamaIndex, CrewAI integrations. See the parent /best/vector-databases guide for non-RAG picks excluded from this lens.

We don't claim "30,000 hours of testing." Our methodology is the formula above plus the editor's published verdict for each pick. Verifiable, auditable, and updated when the underlying data changes.

Why trust Subrupt

We're a subscription tracker first, a buying guide second. Every claim on this page is something you can check.

By use case

Best mainstream RAG default

Pinecone

Read the full review →

Try Pinecone

Best RAG Rust-based OSS

Qdrant

Read the full review →

Try Qdrant

Best RAG Postgres-bundled

pgvector (Postgres)

Read the full review →

Try pgvector (Postgres)

Best RAG hybrid search

Weaviate

Read the full review →

Try Weaviate

How to choose your Vector Database for RAG

RAG read-heavy economics versus realtime write-heavy patterns

RAG workloads are read-heavy by design. Embeddings are written once during ingestion and read many times during retrieval. Pricing economics differ across catalog picks based on this read-heavy pattern. Pinecone Serverless prices reads at fractional cents per million queries; this favors continuous read-heavy production RAG. Qdrant OSS self-host has zero per-query cost on customer infrastructure; this favors high-volume RAG where read-cost dominates. pgvector on Postgres has zero per-query cost beyond Postgres infrastructure; this favors RAG bundled with relational data. Weaviate Cloud Standard prices per dimension and vector count; this favors smaller-scale RAG where retrieval quality lift matters more than per-query cost.

Filter-aware indexing matters for multi-tenant RAG

Multi-tenant RAG, time-bound retrieval, and source-filtered queries combine vector similarity with metadata filters. Pinecone, pgvector, and Weaviate all support metadata filters but degrade in latency when filter cardinality is high or selective filters apply pre-vector-search. Qdrant ships filter-aware HNSW indexing that maintains sub-millisecond latency under filtered queries by integrating filter awareness into the index structure rather than applying filters as post-processing. For RAG with high-cardinality metadata or strict latency budgets under filters, Qdrant is the right call.

Hybrid search: vector plus BM25 keyword retrieval lift

Hybrid retrieval combines vector similarity with BM25 keyword search for retrieval quality lift on data where keyword precision matters as much as semantic similarity. Weaviate ships hybrid search built-in with one query that combines both. Pinecone, Qdrant, and pgvector require application-level merging of separate vector and keyword searches, which complicates the application code and re-ranking logic. The decision pivots on data type. Legal docs, technical documentation, and structured chunks benefit from hybrid retrieval. Pure prose where semantic similarity dominates retrieval quality often runs fine on vector-only retrieval.

When does pgvector beat dedicated RAG vector DBs?

pgvector beats dedicated RAG vector DBs when the team already runs Postgres and the relational data model carries metadata or related entities that RAG retrieval needs. Most production RAG under ten million vectors runs fine on pgvector with HNSW indexing. The break-even depends on team operational capacity. Teams with mature Postgres expertise benefit from pgvector indefinitely. Teams with high-cardinality metadata filtering, extreme query throughput, or hybrid retrieval requirements hit Qdrant filter-aware HNSW or Weaviate hybrid search at lower scale than vanilla pgvector handles. The decision pivots on whether RAG retrieval needs relational joins (pgvector) or vector-specific features (dedicated DBs).

When to look beyond RAG-fit picks (cross-link to parent)

Three patterns push RAG teams beyond RAG-fit picks. First, prototyping in Jupyter notebooks where production scale is not yet load-bearing; Chroma OSS embedded fits prototyping iteration. Second, distributed multi-shard scale past one billion vectors where dedicated distributed architecture matters; Milvus ships GPU-accelerated distributed scale. Third, offline batch analytics on embeddings rather than continuous query RAG; LanceDB ships columnar Lance format with S3 backend for batch workflows. See [our /best/vector-databases guide](/best/vector-databases) for the full lineup including Chroma, Milvus, and LanceDB excluded from this RAG-fit lens.

Frequently asked questions

What is the cheapest vector database for RAG?

It depends on workload posture. OSS self-host on Qdrant, pgvector, or Weaviate has zero licensing cost on customer infrastructure with infrastructure-only spend at $30 to $50 monthly on a small VPS. Pinecone Starter free covers small RAG under 2M reads monthly. pgvector on Supabase Free covers small RAG with relational data. For low-cost production RAG with metadata filtering, self-hosted Qdrant on a VPS is typically the cheapest path.

Should I use Pinecone or self-host Qdrant for production RAG?

Pinecone Standard is closed-source SaaS with serverless pay-as-you-go. Qdrant OSS is Apache 2.0 self-host with unlimited scale on customer infrastructure. The decision pivots on operational preference and metadata filtering needs. Teams without SRE capacity pick Pinecone Standard for the zero-operations path. Teams with mature Kubernetes and metadata-heavy filtering pick Qdrant OSS for the unlimited-scale path with sub-millisecond filtered query latency.

Can I use pgvector for production RAG?

Yes for most workloads under 50M vectors. pgvector with HNSW indexing handles production RAG up to around 50M vectors comfortably on a single Postgres instance. Beyond that scale, query latency degrades and dedicated vector DBs deliver better performance. For teams running RAG against existing relational data, pgvector typically wins on integration cost and operational simplicity at typical SMB and mid-market scale.

Does Subrupt earn a commission from these RAG picks?

On most. We disclose this on every /best page. Free OSS tiers themselves have no transaction. Paid tiers on Pinecone, Qdrant, Weaviate, pgvector via Neon all have plans where we earn commission only on conversion. The composite ranking weights price at 40 percent, features at 30, free tier at 15, fit at 15; none tuned by affiliate rate.

Why is Pinecone ranked first over the cheaper Qdrant?

Pinecone wins on mainstream brand recognition for production RAG because the LangChain, LlamaIndex, CrewAI integrations ship Pinecone as the default during initial RAG selection. Qdrant is genuinely cheaper at scale but the brand-recognition wedge Pinecone holds matches what most production teams default to during initial DB selection. Teams that prioritize cost or OSS posture lean Qdrant; teams that prioritize default-path adoption lean Pinecone.

How does hybrid search improve RAG retrieval quality?

Hybrid retrieval combines vector similarity with BM25 keyword search to capture both semantic similarity and keyword precision. On mixed text-and-metadata data like legal docs or technical documentation, pure vector retrieval misses chunks that contain exact keyword matches but lack semantic similarity. Hybrid retrieval typically improves recall by 10 to 30 percent on datasets with high keyword density. Weaviate ships hybrid built-in; other catalog picks require application-level merging.

What about ChromaDB or Milvus for RAG?

Chroma is excellent for prototyping and small RAG production but limited beyond around 10M vectors due to single-node architecture. Milvus is excellent for distributed-scale RAG past 1B vectors with GPU acceleration but operationally complex for typical SMB RAG workloads. From our catalog the four picks listed here cover the most common RAG paths; readers building prototypes pick Chroma and readers running extreme scale pick Milvus from the parent guide.

How do I evaluate RAG retrieval quality across vector DBs?

Use a fixed evaluation set with ground-truth chunks per query. Measure recall at K (retrieved chunks containing ground-truth chunks within top K) and MRR (mean reciprocal rank) across the evaluation set. Run the same evaluation against each vector DB candidate with the same embedding model and same K value. The differences come from indexing strategy (HNSW vs IVFFlat), filter support, and hybrid retrieval. Plan a 30-day evaluation period running candidates side-by-side before committing.

EU data residency: which RAG picks store data in the EU?

OSS self-host options for Qdrant, pgvector, and Weaviate ship full data residency control through customer infrastructure. Pinecone has multi-region with EU on Standard and Enterprise tiers. Qdrant Cloud has multi-region with EU on Standard and above. pgvector on Neon ships EU regions natively. Weaviate Cloud has multi-region with EU on Standard and above. For default EU residency, OSS self-host on EU infrastructure or Qdrant Cloud EU region are the cleanest catalog fits.

How often is this guide updated?

We re-review pricing and features annually at minimum, with mid-year refreshes when major vendor announcements happen. Pinecone shipped Standard Serverless in 2024. Qdrant Free Cluster launched 2023 and remains always-free. pgvector has shipped consistent Apache 2.0 since 2021. Weaviate Cloud Standard at twenty-five dollars monthly since 2023. The lastReviewed date reflects the most recent editorial pass.

Subrupt Editorial

The team behind subrupt.com. We track subscriptions, surface cheaper alternatives, and publish buying guides where the score formula is on the page so you can recompute it yourself. We do not claim 30,000 hours of testing. What we claim is live pricing from our database, a transparent composite score, and honest savings math against a category baseline.

Last reviewed

Citations

Affiliate disclosure: Subrupt earns a commission when you switch to a service through our recommendation links. This never changes the price you pay. We only recommend services where there's a real cost or feature advantage for you, and our picks are based on the data on this page, not on which programs pay the most.

Related buying guides

Buying guide

Best Threat Intelligence Platforms of 2026

Read guide

Buying guide

Best VPNs of 2026

Read guide

Buying guide

Best Free VPNs of 2026

Read guide

Track your subscriptions on Subrupt

Add the Vector Database for RAG you pay for and see how much you'd save by switching.

Open dashboard

More buying guides

Independent rankings for the subscriptions worth paying for.

See all guides