Salt Technologies AI AI

Vector Database Benchmark 2026

10 databases, 19 fields. Compare query latency, scalability, features, and pricing for RAG and AI search. Pinecone, Qdrant, Weaviate, Milvus, pgvector, and more.

Download CSV View Methodology
Dataset Overview

Records

10

Fields

19

Format

CSV, JSON

License

CC BY 4.0

Version

Q1 2026

Updated

2026-02-15

Publisher

Salt Technologies AI

Source: Salt Technologies AI

CSV JSON

Key Findings

Performance highlights from standardized benchmarks across 1M vectors with 1536 dimensions.

Fastest Latency

Qdrant

4ms p50 query latency

Open source (Apache 2.0)

Best Managed

Pinecone

8ms p50, zero-ops

Serverless scaling

Most Index Types

Milvus

8 algorithms incl. GPU

6ms p50 latency

ACID + Vectors

pgvector

Free PostgreSQL extension

Full SQL + transactions

Query Latency Comparison

p50 and p99 query latency across all 10 databases. Sorted by p50, lowest first. All figures in milliseconds.

p50 latency
p99 latency
milliseconds · 1M vectors, 1536 dim
Qdrant OSS Managed
4ms / 25ms
Redis OSS Managed
5ms / 20ms
Milvus OSS Managed
6ms / 35ms
Pinecone Managed
8ms / 45ms
ChromaDB OSS
12ms / 70ms
Weaviate OSS Managed
12ms / 65ms
Elasticsearch OSS Managed
15ms / 75ms
pgvector OSS
18ms / 90ms
Supabase OSS Managed
20ms / 95ms
MongoDB Atlas Managed
22ms / 110ms
0ms 110ms

Performance and Scalability

All 10 databases with latency, throughput, scaling limits, and deployment options. Scroll horizontally on mobile.

Vector database performance and scalability benchmark Q1 2026
Database Vendor Deployment p50 (ms) p99 (ms) Throughput Max Vectors Max Dims OSS Managed
ChromaDB Chroma Self-hosted (open source, in-process or client-server) 12 70 2,000-8,000 <1M (single node) 65,536 Yes No
Elasticsearch Elastic Managed (Elastic Cloud) + self-hosted (SSPL license) 15 75 5,000-15,000 Billions (distributed) 4,096 Yes Yes
Milvus Zilliz Managed (Zilliz Cloud) + self-hosted (Milvus, Apache 2.0) 6 35 10,000-30,000 Billions+ (distributed) 32,768 Yes Yes
MongoDB Atlas MongoDB Managed (MongoDB Atlas only) 22 110 3,000-10,000 Billions (Atlas) 4,096 No Yes
Pinecone Pinecone Managed (serverless + pod-based) 8 45 5,000-15,000 Billions 20,000 No Yes
pgvector PostgreSQL Community Self-hosted (PostgreSQL extension). Available on managed PG services 18 90 1,000-5,000 10-50M (single node) 16,000 Yes No
Qdrant Qdrant Managed (Qdrant Cloud) + self-hosted (Apache 2.0) 4 25 8,000-20,000 Billions (distributed) 65,536 Yes Yes
Redis Redis Ltd. Managed (Redis Cloud) + self-hosted (Redis Stack, RSAL/SSPL) 5 20 15,000-40,000 10-100M (RAM-bound) 32,768 Yes Yes
Supabase Supabase Managed (Supabase Cloud) + self-hosted. Uses pgvector 20 95 1,000-5,000 10-50M (plan-dependent) 16,000 Yes Yes
Weaviate Weaviate B.V. Managed (Weaviate Cloud) + self-hosted (BSD 3-Clause) 12 65 3,000-10,000 Billions (managed) 65,535 Yes Yes

p50/p99: Median and 99th percentile query latency in ms

Throughput: Vectors indexed per second (bulk insert)

Max Dims: Maximum supported embedding dimensions

Version: Q1 2026 · Last updated: 2026-02-15 · License: CC BY 4.0

Features and Compatibility

Search capabilities, compliance, and SDK support across all 10 databases.

Vector database features and compatibility as of Q1 2026
Database Hybrid Search Multi-Tenancy ACID Index Types Distance Metrics SDK Languages
ChromaDB No No No HNSW cosine, L2, inner product Python, TypeScript, REST
Elasticsearch Yes Yes No HNSW cosine, dotProduct, L2 Python, Node.js, Java, Go, Ruby, PHP, Rust, .NET, REST
Milvus Yes Yes No HNSW, IVF_FLAT, IVF_SQ8, IVF_PQ, SCANN, DiskANN, GPU_IVF_FLAT, GPU_IVF_PQ cosine, L2, inner product Python, Node.js, Go, Java, C#, REST, gRPC
MongoDB Atlas Yes Yes No HNSW cosine, euclidean, dotProduct Python, Node.js, Java, Go, C#, Ruby, PHP, Rust, Swift, Kotlin, REST
Pinecone Yes Yes No Proprietary cosine, dotProduct, euclidean Python, Node.js, Java, Go, REST
pgvector Yes Yes Yes HNSW, IVFFlat cosine, L2, inner product Python, Node.js, Ruby, Go, Java, C#, PHP, Any SQL client
Qdrant Yes Yes No HNSW cosine, dot, euclidean, manhattan Python, TypeScript, Rust, Go, Java, .NET, REST, gRPC
Redis Yes No No HNSW, FLAT cosine, L2, inner product Python, Node.js, Java, Go, C#, Ruby, REST
Supabase Yes Yes Yes HNSW, IVFFlat cosine, L2, inner product TypeScript, Python, Dart, Swift, Kotlin, C#, REST
Weaviate Yes Yes No HNSW, flat, dynamic cosine, dot, L2-squared, hamming, manhattan Python, TypeScript, Go, Java, REST, GraphQL

Pricing Overview

Managed service starting prices and pricing models. All self-hosted open-source options are free.

Redis

$7/mo

Free tier (Redis Cloud, 30MB), pay-per-use from $7/mo. Self-hosted: free (RSAL/SSPL)

Qdrant

$9/mo

Free tier (1GB managed), usage-based from $9/mo. Self-hosted: free (Apache 2.0)

Supabase

$25/mo

Free tier (500MB), Pro at $25/mo, Team at $599/mo. pgvector included

Weaviate

$25/mo

Free tier (managed), pay-per-use scaling. Self-hosted: free (BSD 3-Clause)

MongoDB Atlas

$57/mo

Free tier (512MB), dedicated from $57/mo. Vector search included in Atlas pricing

Milvus

$65/mo

Free tier (Zilliz Cloud), capacity-based from $65/mo. Self-hosted: free (Apache 2.0)

Pinecone

$70/mo

Free tier (2GB), serverless pay-per-use or pod-based from $70/mo

Elasticsearch

$95/mo

Free trial (Elastic Cloud), usage-based from $95/mo. Self-hosted: free (SSPL)

Free self-hosted options

ChromaDBElasticsearchMilvuspgvectorQdrantRedisSupabaseWeaviate

Choosing the Right Vector Database

The right choice depends on your team, scale, and existing infrastructure.

Already using PostgreSQL?

Start with pgvector. Adds vector search to your existing database with zero new infrastructure. Works well up to 10M vectors. For managed hosting with auth and edge functions, use Supabase.

pgvector Supabase

Want zero infrastructure management?

Choose Pinecone. Fully managed, serverless, and no ops overhead. You focus on your application while Pinecone handles scaling, backups, and availability.

Pinecone

Need both vector and keyword search?

Weaviate combines vector similarity with BM25 keyword matching in a single query. Elasticsearch is ideal if you already use it for full-text search.

Weaviate Elasticsearch

Need maximum performance at scale?

Qdrant has the lowest latency in our benchmarks (4ms p50). Milvus offers GPU acceleration and the most index algorithms for tuning performance.

Qdrant Milvus

Need ultra-low latency for real-time apps?

Redis is in-memory with 5ms p50 latency and the highest indexing throughput (15,000 to 40,000 vectors/sec). Ideal for session-based search and caching.

Redis

Already using MongoDB?

MongoDB Atlas Vector Search adds vector capabilities to your existing MongoDB data. No new database to manage. Works within the Atlas aggregation pipeline.

MongoDB Atlas

Data Dictionary

Schema documentation for all 19 fields in this dataset.

Data dictionary for Vector Database Performance Benchmark 2026
Field Type Description Example
database string Official name of the vector database. Qdrant
vendor string Company or organization that develops the database. Qdrant
openSource boolean Whether source code is publicly available and self-hostable for free. true
managedOption boolean Whether a managed/hosted cloud service is available. true
deployment string Available deployment options (managed, self-hosted, or both). Managed + self-hosted
latencyP50Ms number (ms) Median (p50) query latency in milliseconds at 1M vectors. 4
latencyP99Ms number (ms) 99th percentile (p99) query latency in milliseconds at 1M vectors. 25
indexingThroughput string Vectors indexed per second during bulk insert of 100K vectors. 8,000-20,000
maxVectors string Maximum number of vectors supported (depends on deployment). Billions (distributed)
maxDimensions number Maximum supported embedding dimensions. 65536
indexTypes string[] Supported index algorithms (HNSW, IVF, etc.). ["HNSW"]
distanceMetrics string[] Supported distance/similarity metrics. ["cosine", "dot", "euclidean"]
hybridSearch boolean Whether combined vector + keyword search is supported. true
multiTenancy boolean Whether built-in multi-tenant isolation is supported. true
acidCompliant boolean Whether ACID transactional guarantees are provided for vector operations. false
sdkLanguages string[] Officially maintained client SDK languages. ["Python", "TypeScript", "Go"]
managedStartingPriceMonth number | null Starting monthly price (USD) for managed/cloud tier. Null if no managed option. 9
pricingModel string Description of pricing structure and tiers. Free tier, usage-based
bestFor string Recommended use cases based on production experience. Performance-critical RAG

Methodology

How these benchmarks were conducted and validated.

Performance benchmarks were conducted using standardized test conditions: 1 million vectors with 1536 dimensions (OpenAI text-embedding-ada-002), batch sizes of 100 for indexing, and single-query latency measurements at p50 and p99 percentiles. All managed services were tested on their standard tier configurations from US-East regions. Self-hosted databases were tested on AWS r6g.xlarge instances (4 vCPU, 32 GB RAM). Indexing throughput was measured as vectors indexed per second during bulk insert of 100,000 vectors. Filtering benchmarks used metadata filters with 10 distinct filter values. Maximum dimension limits reflect documented specifications as of February 2026. Pricing reflects publicly listed plans; enterprise and committed-use discounts are excluded. SDK language support reflects officially maintained client libraries. Results represent the median of 3 independent test runs. Production experience assessments are based on 50+ RAG deployments across Salt Technologies AI projects in SaaS, healthcare, fintech, and e-commerce.

Cite This Dataset

This dataset is published under the CC BY 4.0 license. Use the citations below to attribute the data in your research, reports, or content.

APA

Salt Technologies AI. (2026). Vector Database Performance Benchmark 2026 (Version Q1 2026) [Dataset]. https://www.salttechno.ai/datasets/vector-database-performance-benchmark-2026/

BibTeX

@misc{salttechnoai_vector_database_performance_benchmark_2026_2026,
  title     = {Vector Database Performance Benchmark 2026},
  author    = {Salt Technologies AI},
  year      = {2026},
  version   = {Q1 2026},
  url       = {https://www.salttechno.ai/datasets/vector-database-performance-benchmark-2026/},
  note      = {Licensed under CC BY 4.0}
}

Version History

This dataset is updated quarterly. All previous versions are documented below.

Version history for Vector Database Performance Benchmark 2026
Version Date Changes
Q1 2026 2026-02-15 Initial release with 10 vector databases across 19 fields including numeric latency benchmarks, indexing throughput, dimension limits, feature flags, and pricing.

Build Your RAG System with the Right Database

Salt Technologies AI selects, configures, and deploys the right vector database for every RAG project.

RAG Knowledge Base

Starting at $15,000 · 3-4 weeks

AI Chatbot Build

Starting at $12,000 · 2-4 weeks

AI Integration Sprint

Starting at $15,000 · 4 weeks

14+

Years of Experience

800+

Projects Delivered

100+

Engineers

4.9★

Clutch Rating

Frequently Asked Questions

Which vector database is best for RAG applications in 2026?
For most RAG applications, Pinecone or Qdrant are the top choices. Pinecone offers the smoothest managed experience with sub-10ms p50 latency and requires no infrastructure management. Qdrant provides the best latency (4ms p50) with both managed and self-hosted options. For teams already using PostgreSQL, pgvector or Supabase is the fastest path to production with minimal infrastructure changes. For applications requiring hybrid search (vector plus keyword), Weaviate combines both in a single query.
Which vector database has the lowest query latency?
Qdrant has the lowest p50 latency among purpose-built vector databases at 4ms, with p99 at 25ms. Redis achieves 5ms p50 due to its in-memory architecture, but is RAM-bound and practical up to 10 to 100 million vectors. Milvus follows at 6ms p50 with GPU acceleration support for even faster performance on large datasets. Pinecone delivers 8ms p50 as a fully managed service with no infrastructure to operate.
What is the cheapest vector database for production use?
pgvector is the cheapest option for teams already running PostgreSQL, adding vector search at zero additional cost. ChromaDB is free and open source for prototyping and small-scale use. Supabase includes pgvector in its free tier with managed hosting. For managed services, Redis Cloud starts at $7 per month and Qdrant Cloud starts at $9 per month. Pinecone and Milvus (Zilliz Cloud) offer higher scalability at higher price points.
Which vector databases support hybrid search?
Weaviate has the most mature built-in hybrid search, combining vector similarity with BM25 keyword matching in a single query. Elasticsearch natively combines kNN vector search with its industry-leading full-text search. Redis supports hybrid search through RediSearch. Qdrant supports hybrid search via sparse vectors. Pinecone offers sparse-dense hybrid search. pgvector and Supabase can combine vector search with PostgreSQL full-text search using standard SQL.
Can I use pgvector or Supabase instead of a dedicated vector database?
Yes, for many use cases. pgvector works well for RAG applications with up to 10 to 50 million vectors and moderate query volumes. Supabase provides managed pgvector with built-in auth, row-level security, and edge functions. Both offer ACID compliance, full SQL, and join support that dedicated vector databases lack. However, purpose-built databases like Qdrant and Milvus offer better performance at scale, more advanced indexing algorithms, and lower latency for high-throughput workloads.
Which vector database scales to the most vectors?
Pinecone, Weaviate, Milvus, Elasticsearch, and MongoDB Atlas all scale to billions of vectors on their managed platforms. Qdrant handles billions in distributed mode. Redis is practical for 10 to 100 million vectors, limited by available RAM. pgvector and Supabase handle 10 to 50 million vectors on a single node. ChromaDB is designed for datasets under 1 million vectors.
Which vector databases are ACID compliant?
pgvector and Supabase are ACID compliant through PostgreSQL, providing full transactional guarantees, rollback support, and consistency for vector operations alongside relational data. No purpose-built vector database (Pinecone, Qdrant, Weaviate, Milvus, ChromaDB) offers full ACID compliance. If your application requires transactional consistency between vector data and relational data, pgvector or Supabase is the recommended choice.
How often is this benchmark updated?
The Vector Database Performance Benchmark is updated quarterly. The current version is Q1 2026, last updated February 2026. We re-run all benchmarks each quarter to reflect new database versions, features, and pricing changes. All previous versions are documented in the changelog.

Need help choosing a vector database?

Start with a $3,000 AI Readiness Audit. We will evaluate your data and recommend the right stack.

Book a Call View All Services