There is no single best vector database. Weaviate leads on out-of-the-box recall, Qdrant on latency, pgvector on simplicity if you already run Postgres, and Pinecone on zero-ops. The right pick depends on corpus size, filters, write pattern and who is on call. This guide compares the options that matter for RAG and AI products, with published benchmarks and pricing, so you can shortlist two and test them on your own data.
TL;DR
- Under a few million vectors and already on Postgres: start with pgvector. Move to a dedicated store only when filters, scale or latency force you to.
- Want no infrastructure to run: Pinecone is fully managed. Want open source you can self-host: Qdrant, Weaviate or Milvus.
- Published benchmarks disagree because conditions differ. Judge on your own corpus, filters, concurrency and p99 latency.
- In regulated work (finance, healthcare, legal), data control and tenant isolation matter as much as speed.
- We pick the retrieval layer as part of a full AI-native build, so the database choice follows your compliance needs.
Table of Contents
- What a vector database actually has to do
- The best vector databases at a glance
- Pinecone: the managed default
- Weaviate: hybrid search and built-in modules
- Qdrant: latency and filtering
- Milvus: built for scale
- pgvector: when Postgres is enough
- Redis, Turbopuffer, Chroma, LanceDB and FAISS
- Pinecone vs Weaviate, Milvus vs Qdrant
- Why benchmarks disagree
- How to choose: five questions
- What about the AWS-native options?
- Our perspective
- Build the whole system, not just the index
- FAQ
- Sources
What a vector database actually has to do
A vector database stores embeddings and returns the nearest ones to a query fast. For RAG, that is only the start. One hands-on comparison organizes selection around three pillars: retrieval quality (ANN algorithms such as HNSW or IVF, recall tuning, hybrid dense plus sparse search), latency and throughput (sub-100 ms for interactive RAG), and metadata filtering, including how efficiently filters are applied.
A 2026 academic evaluation of seven systems (FAISS, Qdrant, Milvus, Weaviate, Chroma, pgvector and LanceDB) scored them on retrieval quality, latency, throughput, resource use and operational complexity. Use the same five lenses when you shortlist.
The best vector databases at a glance
| Database | Model | Strongest for | Entry pricing (published) |
|---|---|---|---|
| Pinecone | Fully managed, proprietary | Teams without a platform group | Starter $0, Builder $20/month, Standard from $50/month |
| Weaviate | Open source or managed cloud | Hybrid search, built-in vectorization | Free $0, Flex from $45/month |
| Qdrant | Open source (Rust), cloud or self-hosted | Low latency, selective filters | Free cloud tier, single node |
| Milvus | Open source, Zilliz Cloud managed | Very large scale | Zilliz Free, dedicated from $126/month |
| pgvector | Postgres extension | Teams already on Postgres | Your Postgres cost |
| Redis | In-memory platform with vector search | Speed, unified cache and vectors | Varies by deployment |
| Turbopuffer | Commercial, object-storage based | Cost-efficient scale | Launch $16/month minimum |
| Chroma, LanceDB | Lightweight, embedded | Prototyping, local development | Open source |
Each row is expanded below with what the evidence supports.
Pinecone: the managed default
Pinecone is fully managed and proprietary, with no self-hosting option. It offers dense and sparse indexes, full-text search, metadata filtering and reranker support, according to a comparison of the two products.
On its pricing page, Starter is free, Builder is $20/month, Standard has a $50/month minimum and Enterprise has a $500/month minimum. Enterprise includes BYOC, private endpoints and HIPAA compliance. Another pricing review checked on 1 November 2026 lists a 99.95% uptime SLA on Enterprise and a free tier with 2 GB of storage.
Pinecone fits bursty traffic when nobody wants to run a stateful service. The trade is that your vectors live in its cloud, so check that its deployment options match your data-residency rules.
Weaviate: hybrid search and built-in modules
Weaviate is open source under a BSD-3 license and supports HNSW, Flat and Dynamic indexes with native hybrid search. Redis’s comparison lists its built-in embedding and rerank modules as a best fit. You can self-host with Docker, Kubernetes or bare metal, or use the managed cloud.
A pricing review checked on 1 November 2026 shows Weaviate Cloud Free at $0 (100,000 objects), Flex from $45/month with a 99.5% uptime SLA, and Premium from $400/month prepaid with a 99.95% SLA. HIPAA is available on Enterprise Cloud (AWS).
In the academic evaluation, Weaviate had the best out-of-the-box recall, above 99% on SIFT1M. In a filtered, multi-tenant RAG scenario in the AIMultiple benchmark, it held 1.00 filtered recall, and it reached 8,330 QPS with a 32-worker client.
Qdrant: latency and filtering
Qdrant is an open-source Rust database offered as Cloud, Hybrid Cloud, Private Cloud and Edge (Beta). Its free cloud tier is a single node with 0.5 vCPU, 1 GB RAM and 4 GB disk. It supports native dense and sparse hybrid search, one-stage filtering and quantization that can cut memory by up to 64x, per a third-party comparison with Milvus.
In the arXiv evaluation, Qdrant gave the best median latency among full databases (4.55 ms) and led single-thread throughput among them on SIFT1M at 216 QPS. That makes it a strong candidate when filters are selective and you can run a stateful service.
Milvus: built for scale
Milvus is an open-source Go and C++ database with Lite, Standalone and Distributed modes, plus managed Zilliz Cloud. It is designed for deployments scaling to tens of billions of vectors, according to the same comparison. In the academic study it performed best on the 960-dimension dataset, and in the AIMultiple filtered scenario it held 0.984 recall at up to 732 filtered QPS.
On pricing, a Zilliz Cloud review from July 2026 lists a Free tier with 5 GB storage, Standard serverless from $0/month or dedicated from $126/month, and Enterprise with audit logs, SSO and granular RBAC.
Milvus makes sense when you expect very large corpora and have engineers to operate the distributed mode, or you are happy to pay for Zilliz Cloud.
pgvector: when Postgres is enough
The academic evaluation highlights pgvector for SQL integration, and its single-thread throughput on SIFT1M (154 QPS) landed second among full databases. Quality differences are small: at a fixed Recall@10 of 0.95 on MedRAG-50k, nDCG@10 ranged only from 0.803 for pgvector to 0.817 for LanceDB.
One production rule of thumb says that under a few million vectors with Postgres already in production, use pgvector with HNSW and stop there. That is sensible. You avoid a second system, a second backup plan and a second access-control model.
In regulated settings, tenant isolation deserves care. One engineering write-up warns that if tenant filtering is enforced only in application code, a missed code path can leak results across tenants. It recommends database-level enforcement such as PostgreSQL row-level security, or dedicated namespaces, for compliance-grade isolation.
Redis, Turbopuffer, Chroma, LanceDB and FAISS
Redis offers HNSW, FLAT and SVS-VAMANA indexes and hybrid search through FT.HYBRID, according to its own comparison. That article argues a unified platform means fewer systems and no network hop between vector search and a cache. In the AIMultiple test, Redis delivered the highest single-thread throughput (764 QPS) at matched recall.
Turbopuffer is commercial only. Its pricing page lists a Launch plan with a $16/month minimum and Enterprise from $4,096/month with a 35% usage premium, including single-tenancy, BYOC and a 99.95% uptime SLA.
Chroma and LanceDB suit prototypes and embedded use. LanceDB trades some recall for faster indexing in the academic evaluation, and it was slowest in the AIMultiple single-thread test (70 QPS).
FAISS is a library rather than a database. It leads throughput in the academic study (866 QPS) but lacks database operational features such as persistence management and access control.
Pinecone vs Weaviate, Milvus vs Qdrant
These are the two matchups readers search most.
Pinecone vs Weaviate. Choose Pinecone if you want fully managed with minimal operations. Choose Weaviate if you want open source, self-hosting or built-in vectorization. Both support hybrid vector and keyword search. Dimension limits differ: one comparison lists up to 20,000 per dense vector for Pinecone and up to 65,535 for Weaviate. Weaviate implements filtered search as pre-filtering. On certifications, the same page lists SOC 2 Type II and ISO 27001:2022 for Pinecone, with a HIPAA BAA on request, and SOC 2 Type II and ISO 27001 for Weaviate.
Milvus vs Qdrant. Choose Milvus for the largest corpora and a choice of deployment sizes. Choose Qdrant for low latency, strong filtering and lighter operations at moderate scale. No neutral head-to-head page gives definitive numbers for this pair, so test both on your data.
Why benchmarks disagree
Results depend on conditions. In the AIMultiple run, Redis won single-thread throughput while Weaviate won under a 32-worker client. In the academic run, Qdrant led single-thread throughput among full databases. A community repository listing 100M-vector results shows Weaviate with higher peak QPS than Qdrant, and it is one person’s test setup, so treat it as one data point.
A production-minded critique lists the usual flaws: frozen datasets, no concurrent writes, mismatched hardware and filters so selective they discard most data. Another source argues that for agents, concurrent p99 latency determines the SLA, not serial latency or peak QPS.
How to choose: five questions
- How big is the corpus, and how many dimensions? A few million vectors rarely need a distributed system.
- How selective are your filters? Selective filters favor Qdrant, Milvus or Weaviate; one guide groups these with Elasticsearch for teams willing to run a stateful service.
- Is the write pattern bulk or continuous? Continuous upserts and deletes change which index holds up.
- What is your p99 target under concurrent load? Measure it, not the average.
- Who is on call? If nobody wants to be, a managed service is cheaper than it looks.
The same source advises benchmarking candidates on your own corpus and filters. Plan for a short bake-off of two candidates with real queries.
Regulated data: control and isolation
If you handle patient records, client files or financial documents, add three requirements: where the vectors physically live, how tenants are isolated, and whether audit trails exist. Cloud tiers differ here. Pinecone’s Enterprise plan lists BYOC, private endpoints and HIPAA, and Weaviate lists HIPAA on its Enterprise Cloud (AWS). Self-hosting an open-source engine inside your own environment is the other route.
If you want to see how retrieval connects to model choice, our guide on RAG versus fine-tuning for regulated teams covers that decision. Filtered, metadata-driven retrieval also shapes how content gets found by AI search, a topic explored in vector search for AI citations.
What about the AWS-native options?
If you are all-in on AWS, the official guidance points to different services by need. For ultra-low latency, it recommends Amazon MemoryDB, citing sub-millisecond response times. For billions of vectors with infrequent retrieval, it describes Amazon S3 Vectors with up to 90% cost savings for workloads that tolerate sub-100 ms latency. These are AWS’s own claims, so verify them against your workload.
Our perspective
Most teams treat the vector database as the decision. It is one component in a system, and the system is where the value sits.
We don’t sell AI projects, we invest in firms. Our team comes from regulated industries, including ISO 27001, pharma and finance, where getting AI wrong is not an option. So we start from your constraints: what data may leave your environment, who needs audit trails, and which workflow you want to automate.
Our approach is designed around private and self-hosted LLM deployment so your data stays under your control. The retrieval layer is chosen to serve that, whether that ends up being pgvector inside your existing stack or a dedicated engine. You bring the domain expertise and client base. We automate your delivery and co-build an AI-native product with you that you own.
Matevz
Build the whole system, not just the index
If you run a law, accounting, healthcare, insurance or consulting firm, picking a database is the easy part. Turning your expertise into a working, compliant system is the hard part. At Autonomousfirm we build custom AI-native systems for regulated industries, with compliance, security and IP ownership built in, on a partnership or venture basis. Our platform line includes AI OS, plus Compliance OS, Knowledge OS and more, all described on our main site.
Tell us the workflow you want to automate, and we will show how the retrieval layer fits into a system you own.
FAQ
Which vector database is the best overall?
None wins everywhere. Weaviate had the best out-of-the-box recall in a 2026 academic evaluation, Qdrant the best median latency among full databases, and FAISS the highest throughput but without database features. Match the engine to your corpus, filters and team.
Is pgvector good enough for production?
Often yes, for corpora under a few million vectors when you already run Postgres. You keep one system and can use row-level security for tenant isolation. Move to a dedicated database when filter selectivity, write volume or scale starts to hurt latency.
Which vector database is fastest?
It depends on the test. At matched recall, Redis led single-thread throughput in the AIMultiple benchmark and Weaviate led under 32 concurrent workers. Qdrant led among full databases in the academic study. Measure p99 latency under your own concurrent load.
Are vector databases going away?
Not on this evidence. Many teams now use vector search inside databases they already run, such as Postgres or Redis, instead of a separate product. The capability is staying. What changes is where it lives.
Sources
- We Tried and Tested 10 Best Vector Databases for RAG Pipelines
- A Comprehensive Empirical Evaluation of Vector Database Systems (arXiv)
- Vector Database Benchmark: 7 Open-Source Engines for RAG
- Comparing the best open source vector databases (Redis)
- Pinecone vs Weaviate: Managed vs Open-Source
- Milvus vs Qdrant: Scale or Search Precision?
- Pinecone vs Weaviate (inetgeek)
- Pricing | Pinecone
- turbopuffer pricing
- Pinecone pricing review
- Weaviate pricing review
- Milvus and Zilliz Cloud pricing review
- How to Choose a Vector Database for Production
- Vector Databases: pgvector vs Dedicated Stores
- AWS Prescriptive Guidance: Choosing an AWS vector database for RAG
- vectorbench repository
- Vector Database Benchmarks Are Lying to You
- OceanBase seekdb launch blog
Recommended
- Engineers: NIST Aligned AI Workflow Orchestration That Resumes - benchmarked
- Three Commitments to Scale AI Adoption for Regulated Enterprises - benchmarked
- Avoid 1.5–2x Cost Surprises: AI Project Costs for Procurement - benchmarked
- Regulated executives: AI implementation roadmap with ISO/IEC 42001 - benchmarked
