Which vector database should I use for RAG?
Below a few million vectors, the extension in the database you already run — usually pgvector — is the right answer. A dedicated vector database earns its operational cost at scale, not at the start.
Surveyed 10 August 2026
Add vector search
Open in Tools →- 01Qdrant
Qdrant is a high-performance, open-source vector database and similarity-search engine written in Rust, built for storing and searching embeddings at massive scale for RAG and AI applications.
Developer Tools - 02Chroma
Chroma is an open-source AI-native search and retrieval database designed to be the easiest way to add embeddings-based retrieval to an app, popular as the default vector store in LLM prototypes.
Developer Tools - 03pgvector
pgvector is an open-source Postgres extension that adds vector similarity search to a database you already run — store embeddings in a column and query them with SQL for RAG.
Developer Tools - 04Weaviate
Weaviate is an open-source vector database that combines vector and structured (hybrid) search with cloud-native scalability, storing both objects and their vectors for retrieval and RAG.
Developer Tools - 05Pinecone
Pinecone is a fully-managed, serverless vector database that lets developers store and query embeddings at scale without running infrastructure — a common production retrieval layer for RAG.
Developer Tools - 06LightRAG
Simple, fast retrieval-augmented generation system from HKUDS that combines vector search and graph-based knowledge representation. EMNLP 2025.
AI Tools
Ordered by the VibeLeaderboard index, re-surveyed each edition. Something missing?
What to look for
- 01Do you actually need a separate service? One less system to operate is worth a lot below a few million vectors.
- 02Can you filter by metadata and vector in one query? Splitting them wrecks both recall and latency.
- 03Does it support hybrid search? Pure semantic retrieval misses exact identifiers, names, and error codes.
Common questions
- Is pgvector good enough for production RAG?
- For most applications, yes. With an HNSW index it handles millions of vectors at low latency, and keeping embeddings beside your relational data makes filtered queries straightforward.
- Why does my RAG miss obvious results?
- Usually chunking or pure-semantic retrieval. Chunks that split a fact across boundaries never match well, and embeddings alone are poor at exact terms — add keyword search alongside.
More in Understand the AI landscape
- Find AI benchmarksBrowse the evaluations used to compare models, agents, coding, and cost.
- Evaluate an LLM applicationBuild test sets, score outputs, and catch quality regressions.
- Observe an LLM applicationTrace calls, inspect failures, and monitor latency, quality, and spend.
- Run models locallyUse local inference runtimes and model managers on your own hardware.