Chroma

Chroma keeps documents, metadata, and their vector embeddings together, then answers nearest-neighbor queries so an application can fetch relevant context for an LLM prompt. You can begin with an ephemeral in-memory client inside a Python script, move to a persistent client that writes to a local folder, or launch a standalone server with the chroma CLI and connect from Python or JavaScript over HTTP. Hand it raw text and a configurable embedding function vectorizes it, with adapters for OpenAI, Cohere, Hugging Face, and sentence-transformers, or pass precomputed vectors yourself. Queries blend similarity ranking with metadata conditions, plus keyword and regex matching over stored text, and collections can index images next to text. Most of the engine is now written in Rust, yet setup is still one pip or npm package with no GPU needed. The whole repository is Apache 2.0, so self-hosting inside commercial products is fine; Chroma Cloud is an optional paid managed tier with a newer search API and connectors that sync from S3, GitHub, or websites. Well suited to RAG prototypes and small to mid-size production retrieval.