pgvector alternatives: 6 tools compared (October 2026)

turbopuffer is a serverless search engine that keeps all data in object storage and caches it on NVMe SSD and memory for queries. It serves vector, full-text and hybrid search for AI applications with many namespaces and large document counts.

Key functions

• Approximate vector search with dense and sparse vectors, combinable with filters
• BM25 full-text search and hybrid search
• Object storage as the only source of truth, with caching for fast queries

Compared with pgvector

• Publishes a price (from $16 per month); pgvector quotes on request.
• Can store data in the EU; pgvector states no EU option.
• Built for larger organizations than pgvector's smallest customers.

Check before buying: Check whether higher write latency and occasional cold queries fit your application, as the vendor documents these tradeoffs. Choose the region explicitly for EU data. Single sign-on and audit logs start at the scale tier, and private networking and CMEK at enterprise. A SOC 2 report and a DPA are listed on all tiers.

Weaviate is an open-source vector database from Amsterdam that stores objects together with their vectors. Teams use it for semantic and hybrid search, retrieval-augmented generation and agent workflows, self-hosted or in Weaviate Cloud.

Key functions

• Semantic and hybrid search across objects and vectors
• Hosted embedding models billed per million tokens
• Query Agent turns natural-language questions into database operations

Compared with pgvector

• Publishes a price (from $45 per month); pgvector quotes on request.
• Strongest elsewhere: Netherlands.

Check before buying: Check which regions your plan offers, since Flex is limited to a smaller set of regions than Premium. SSO, SAML and HIPAA compliance are listed for the Premium tiers only. Vector dimension prices vary by provider, region and compression, so use the cost module in the console before committing. Request the data processing agreement from Weaviate B.V.

Zilliz offers Zilliz Cloud, a managed vector database built on the open-source Milvus engine. Development teams use it to store embeddings and run similarity, keyword and hybrid search behind RAG applications, recommendations and semantic search, on AWS, Google Cloud or Azure.

Key functions

• Vector, filtered, full-text and hybrid search with reranking for RAG and semantic search
• Three cluster types for real-time serving, large datasets or low-cost tiered storage
• On-demand compute for search and indexing jobs on external data in a data lake (public preview)

Compared with pgvector

• Publishes a price (from $197 per month); pgvector quotes on request.
• Can store data in the EU; pgvector states no EU option.
• Names 2 security certificates, pgvector 0.

Check before buying: Check cost drivers early: Dedicated clusters price by cluster type and vector count, Serverless by usage, so model your data volume with the vendor's calculator. SSO, audit logs and private endpoints start with the Enterprise plan, and the vendor lists HIPAA readiness under Business Critical. US privacy rules for stored personal data, such as state laws and HIPAA, remain the buyer's task. The vendor publishes a sub-processor list in its trust center.

Chroma is an open-source embedding database for building retrieval into AI applications. Developers add documents, and Chroma embeds, indexes and searches them through a small API, locally or in the managed Chroma Cloud.

Key functions

• Handles tokenization, embedding and indexing when documents are added
• Similarity queries with filters on metadata and document content
• Chroma Cloud with serverless vector, hybrid and full-text search

Compared with pgvector

• Publishes a price; pgvector quotes on request.
• Can store data in the EU; pgvector states no EU option.
• Strongest elsewhere: United States.

Check before buying: If you self-host, plan persistence, backups and upgrades yourself, since the project releases new versions weekly. For Chroma Cloud, choose the EU region when data must stay in Europe and check the usage-based costs for writes, storage and queries. SOC 2 is listed for the Team plan, and single-tenant or BYOC clusters require Enterprise. Chroma publishes a data processing addendum and a sub-processor list for its cloud service.

BGE-M3 is an open embedding model for search and retrieval-augmented generation (RAG) that supports more than 100 languages and three retrieval methods, widely used as a building block for company knowledge search.

Key functions

• Dense, sparse and multi-vector retrieval in one model
• More than 100 languages
• Inputs of up to 8,192 tokens

Compared with pgvector

• Strongest elsewhere: China.

Check before buying: Self-hosting means your team runs, secures and updates the model, including GPU capacity, monitoring and patching. Check the license terms for commercial use and keep a record of which version you deploy. If outputs feed decisions about people, a human must stay in the loop. Changing the embedding model later means re-embedding all documents, so test retrieval quality on your own data first.

LanceDB is an open-source, embedded retrieval library and database for multimodal AI data. It stores vectors together with text, images and video in the Lance columnar format and runs locally or in the cloud without a separate server.

Key functions

• Vector search over very large datasets with GPU-assisted index building
• Full-text and SQL search next to vector search
• Stores vectors with images, video and other raw data, with automatic versioning

Compared with pgvector

• Strongest elsewhere: United States.

Check before buying: Decide whether the embedded open-source library fits, or whether you need LanceDB Cloud or Enterprise for managed operation; prices for these are available on request. If you store data on your own object storage, backups and access control stay with you. The Apache 2.0 license allows commercial use.