Chroma

Chroma develops open-source search and retrieval infrastructure for AI applications. Its database supports dense vector (semantic similarity) search, sparse vector and lexical search (BM25, SPLADE), full-text search including trigram and regex, and metadata filtering and faceted search, with additional capabilities for dataset forking (versioning, A/B testing, roll-outs), multi-modal indexing of text, images and audio, and a command-line tool for development. The core project is licensed under Apache 2.0 and can be run locally or self-hosted, with SDKs shown for Python, TypeScript and Rust and support for external embedding providers including OpenAI, Cohere, Hugging Face and sentence-transformers.

The architecture separates a query layer with in-memory (hot) and SSD (warm) caches from a storage layer on object storage such as S3 or GCS holding all vectors, metadata and indexes, with query-aware tiering between them. The company positions this design as substantially cheaper than memory-resident systems, citing object storage at roughly $0.02/GB/month versus about $5/GB/month for memory, and claims up to 10x lower cost. Published figures include p50/p90/p99 query latency of 20/27/57 ms warm and 650 ms/1.2 s/1.5 s cold at 384 dimensions over 100k vectors, write throughput of 30 MB/s (2000+ QPS) per collection, 10 concurrent reads (200+ QPS) per collection, 1M collections per database, 5M records per collection, and 90–100% recall.

Alongside the managed Chroma Cloud service, the company offers an enterprise deployment model with bring-your-own-cloud in the customer's VPC (data plane in the customer's cloud, control plane managed by Chroma), multi-cloud and multi-region replication, and point-in-time recovery, and states SOC 2 Type II compliance. Adjacent products listed include Sync, Package Search MCP and a hosted embedding service. Chroma also publishes research on search, retrieval, agents and context engineering, including Context-1 (a self-editing search agent), Context Rot (effects of increasing input tokens on LLM performance), generative benchmarking for retrieval evaluation, and chunking strategies.

Founding story

Chroma grew out of its founders' work on model interpretability: while experimenting with understanding models through their latent spaces, they needed a powerful and easy-to-use open-source vector database, evaluated existing products and found them hard to use and aimed at a different use case, and built Chroma for themselves. The seed announcement is signed by Jeff and Anton, who are described as having deep interpretability experience.

Business model

Chroma pairs a permissively licensed (Apache 2.0) open-source database that developers can run locally or self-host with commercial hosted offerings. Chroma Cloud is a managed, serverless service with serverless pricing, a free tier and $5 in free credits at sign-up, structured into tiers that include a Pro plan (with direct Slack support) and an Enterprise plan (customized SLAs, BYOC in the customer's VPC).

Usage-based serverless pricing for the managed Chroma Cloud database, plus paid Pro and Enterprise subscription tiers that add support levels, SLAs and enterprise deployment features such as BYOC, multi-region replication and point-in-time recovery.

Traction

The website reports over 15M monthly downloads and 27k GitHub stars, with other page sections citing over 26k stars, use in more than 90k other open-source codebases on GitHub, over 11M monthly downloads, and a Discord community of about 10,000 people. Five weeks after launch in 2023 the company reported crossing 35k Python downloads in a month. Prefect CEO and co-founder Jeremiah Lowin publicly endorsed Chroma's ease of adoption and recommended it to Prefect customers.

Latest developments

Recent product surface includes Chroma Cloud with serverless pricing and a free credit offer, sparse/lexical search (BM25, SPLADE), Schema() and Search() APIs, forking for dataset versioning and A/B testing, a CLI, Package Search MCP, and a hosted embedding service, all reported operational on the company status page. Research output includes Context-1, a self-editing search agent, plus Context Rot, generative benchmarking and chunking-strategy studies.