Chroma

Chroma is an open source search engine for AI applications, licensed Apache 2.0. It stores documents, their embeddings and their metadata in collections and retrieves them by vector similarity, by keyword, or by both combined, with metadata filters applied alongside the search. It runs in three shapes with a consistent API: as an embedded library inside a Python, TypeScript or Rust process for prototyping; as a single server started with one command, intended for small and medium workloads, which the project characterises as fewer than about ten million records across a handful of collections; and as a distributed deployment of independent services over object storage with SSD caches and a shared system database, which is what Chroma Cloud runs. What distinguishes it is the on-ramp. Almost every competing vector database asks you to run something before you can index your first document; Chroma asks for an import statement, and the code you write against it in a notebook is the code that runs against the cloud. Commercially that has made it the default in tutorials, framework integrations and agent toolkits, which is a real advantage because it means the retrieval layer is rarely the part of a prototype that needs explaining. The permissive licence reinforces this: there is no source-available clause to review before embedding it in a product. The buyers are application teams building retrieval-augmented generation and agent memory, usually starting embedded and deciding later whether to move to the server or the cloud. The trade-off is that the easy mode has a physical ceiling. On a single node the amount of system memory bounds collection size directly, roughly a quarter of a million records per gigabyte of RAM at 1024 dimensions, and query throughput parallelises only up to the number of vCPUs before latency rises linearly with concurrency. The distributed deployment is a genuinely different architecture, with different latency and consistency behaviour, so the thing you tested locally is not the thing you eventually run. Planning for that transition early is the difference between a smooth migration and a rewrite under load.