When AI applications need to find 'things like this' among millions or billions of items, vector databases do the finding. They store embeddings — numerical representations of text, images, and other data — and answer similarity queries in milliseconds, which makes them the backbone of RAG pipelines, semantic search, recommendations, and agent memory. Users are AI and platform engineers; the buying decision usually lands with whoever owns retrieval infrastructure.
Under the hood, these systems combine approximate nearest neighbor indexes (like HNSW) with the trappings of real databases: metadata filtering, hybrid vector-plus-keyword search, replication, and horizontal scale. The market splits along familiar lines. Managed-first: Pinecone ($228M raised) built the category's best-known serverless offering, and Zilliz ($113M) runs the managed service behind open-source Milvus. Open-source-first: Weaviate ($50M) and Qdrant ($50M) — the latter built in Rust for performance — plus Chroma, popular as the default for prototyping. Architectural challengers rethink the cost model: Turbopuffer serves vectors directly from object storage for massive scale at low cost, LanceDB positions as a multimodal lakehouse, and Vespa.ai brings decades of big-data serving heritage to combined vector and lexical search.
Leaders separate on recall-versus-latency at scale, filtering performance (a common real-world bottleneck), and cost per million vectors. Before buying, benchmark with your own embeddings and query patterns, decide managed versus self-hosted, and check hybrid search quality — pure vector similarity alone often underperforms in production. Also weigh whether your existing database's vector extension is simply good enough. NeuronFeed tracks 8 vector database companies with $532M in combined funding.