Chroma

Chroma is data infrastructure for AI applications, the place embeddings live between the model that produces them and the application that searches them. Getting started takes one pip install and no server process, because the client runs entirely in memory while you prototype and gains persistence later without a rewrite.

The core API is four functions. You create a collection, add documents to it, query it, and get items back by id, with update and delete available on the same collection object. Tokenization, embedding and indexing are handled for you, though you may supply your own vectors instead.

What that small surface gives you in practice is the following.

• Collections: named groups of documents, created, fetched or deleted through the client.
• Automatic embedding: text is tokenized, embedded and indexed on insert unless you pass vectors yourself.
• Metadata filters: attach metadata to each document and constrain queries with a where clause.
• Document filters: narrow results by matching on the text of the documents themselves.
• Client-server mode: run Chroma against a database path to serve it over the network.

Python and JavaScript clients are published to PyPI and npm, with tagged releases going out on Mondays and hotfixes at any point in the week. Chroma Cloud is the hosted counterpart, serving vector, hybrid and full-text search for teams that would rather not operate the database. The project is Apache 2.0.