Chroma

Chroma is an open-source vector database for AI retrieval, licensed Apache 2.0 to self-host or run as Chroma Cloud, a serverless service billed by usage.

Chroma gives AI applications a place to keep documents with their embeddings and pull back the most relevant ones when a model needs context. Chroma Inc., the San Francisco firm that develops it, has no parent company. One Apache 2.0 codebase powers both the open-source database and Chroma Cloud, the managed service. Its scope reaches past a plain vector store into full-text, regex and metadata search. Each collection holds records made of an embedding, the source document and metadata fields. Single-node Chroma indexes them with HNSW, while Chroma Cloud uses a SPANN index. Distance is set per collection as l2, inner product or cosine. The Search API applies metadata filters inside the ranked query. It also fuses dense and sparse rankings, such as BM25 or SPLADE, with Reciprocal Rank Fusion. Chroma Cloud can embed text on its own servers through hosted Qwen and SPLADE embedding functions. Forking creates a copy-on-write collection instantly for dataset versioning, A/B tests and roll-outs. Chroma Sync parses, chunks and embeds files from S3 buckets, GitHub repositories and websites. Chroma suits developers building retrieval-augmented generation, agents and AI search. A solo developer or a small team can run the single-node engine through pip, npm or Docker, in memory or with persistent storage. Production teams that would rather not operate servers move to Chroma Cloud, which needs no provisioning or tuning. Collections copy between a local install and the cloud with the Chroma CLI. Organizations with security requirements can deploy into their own VPC with the bring-your-own-cloud option, or on single-tenant clusters. Grounding a model's answers in a team's own documents is the job it is hired for most. Engineers usually wire it in through LangChain or LlamaIndex, with embeddings from OpenAI, Cohere or Hugging Face when they bring their own model. Agent builders use it as memory, so an agent can search what it saw earlier before acting. Sync loads a code or documentation corpus without a separate ingestion pipeline. Clients for Python, TypeScript and Rust, plus a command-line tool, carry a project from a notebook to a deployed service.