Chroma Context-1

by Chroma

20B agentic search model that rewrites RAG pipelines — retrieves, validates, and reformulates queries for multi-hop reasoning

See https://huggingface.co/chroma/Context-1

Features

  • 20B parameter agentic search model — purpose-built for retrieval, not general chat
  • RAG pipeline rewriting — reformulates retrieval prompts for higher-quality context injection
  • Multi-hop query support — handles complex queries requiring chaining across multiple documents
  • ChromaDB integration — designed to work natively with the ChromaDB vector store ecosystem
  • Locally deployable — run via Hugging Face model weights on sufficient hardware

Superpowers

Context-1 attacks a specific weak point in typical RAG pipelines: the query reformulation step. Standard RAG sends the user’s raw question to the retriever, which often fails on complex or multi-part questions. Context-1 acts as a dedicated retrieval model that reformulates the query into a retrieval-optimized form, extracts supporting documents, and validates their relevance before passing context to the generation model. This separates retrieval intelligence from generation intelligence — rather than asking one model to do both, you get specialized performance at each stage. Useful for production RAG systems where retrieval quality is the bottleneck.

Pricing

  • Open weights — free to self-host