Chroma

Chroma Cloud is now available via Stripe Projects! Stripe Projects allows your agent to provision infrastructure on your behalf. Projects automatically creates accounts, manages billing and provisions API keys, cutting out the frustration of going into the browser to sign up for services. Get started now! stripe projects add chroma/database

New: EU region support for Chroma Cloud. Create databases in GCP’s europe-west1. Available now on all plans.

Chroma offers a variety of lexical search strategies. Full Text Search (FTS), BM25 and SPLADE are the core offerings. Each have their strengths and weaknesses and use cases. Learn more 👇

We're releasing Chroma Context-1, a 20B parameter open-source search agent that pushes the pareto frontier of agentic search, an order of magnitude faster and cheaper than frontier alternatives. Accurate search is rarely a single step. The output of one search informs the next. Frontier LLMs can do this through agentic search, but long trajectories become cost and latency prohibitive. Context-1 solves this by separating search from generation. Three ideas that made it work: 1. Staged training: recall first, then precision. Context-1 is trained with SFT + RL on 8,000+ synthetic multi-hop tasks. The curriculum first optimizes for broad recall, then progressively trains the agent to narrow down to the most relevant documents. The result is a model that retrieves thoroughly and selects carefully. 2. Self-editing context. As the agent searches, its context window fills with documents, many of which may be irrelevant. Context-1 is trained to selectively prune its own context mid-search, freeing space for further exploration and reducing context rot. This lets a 20B model with a 32k token budget outperform frontier models with much larger context windows. 3. Scalable synthetic task generation. We built an extraction-based verification pipeline with an LLM judge that achieves high human alignment, minimizing the need for manual annotation. Tasks span 4 domains: web, SEC filings, patent law, and email, each requiring the agent to chain clues across documents. Context-1 matches or exceeds frontier models on BrowseComp-Plus, SealQA, FRAMES, HotpotQA, and HLE. We're open-sourcing the model weights, the harness and the full task generation codebase. Apache 2.0. Full report in comments.

Chroma Cloud is launching with Stripe Projects. The high friction points of building with AI isn't the AI, it's everything around it. Adding services means managing account credentials, API keys, user permissions, a new dashboard to navigate and more. Stripe Projects removes this friction: stripe projects add chroma/database The Stripe Projects CLI will provision a Chroma Cloud account, generate and store credentials and manage billing, no browser involved.

Collection Forking on Chroma Cloud unlocks faster workflows on top of your data without the overhead of starting from scratch.

Gemini Embedding 2 is available in Chroma clients today!

$1000 of Chroma Cloud credits are now available within the Fin.ai Startup Pack!

New: Chroma Sync is serverless data ingestion for Chroma Cloud. Automatically ingest: - GitHub repositories - Websites - S3 buckets Data pipelines can be brittle and hard to maintain. Chroma Sync makes ingesting data for search simple.

New: Our agent skill can add search to your product with one prompt.