Model overview
context-1 is a 20 billion parameter search model built by chromadb that specializes in retrieving supporting documents for complex queries that require multiple reasoning steps. Unlike general-purpose language models, this model operates as a specialized agent designed to work alongside frontier reasoning models, handling the retrieval component of agentic systems. It achieves retrieval performance comparable to larger models while running up to 10 times faster and at a fraction of the cost. The model uses a Mixture of Experts architecture and was trained through supervised fine-tuning combined with reinforcement learning using a staged curriculum approach.
Model inputs and outputs
context-1 accepts user queries and operates within an agent harness that manages tool execution and context management. The model processes queries by decomposing them into subqueries, executing searches across document corpora, and managing its own context window through selective pruning. It returns ranked documents relevant to the original query along with intermediate search steps.
Inputs
- Natural language queries: User questions or information requests, including complex multi-hop questions with multiple constraints
- Document corpus: A collection of documents to search across
- Agent harness signals: Context budget information and pruning instructions from the agent framework
Outputs
- Retrieved documents: Relevant documents ranked by relevance to the query
- Tool calls: Formatted requests for document retrieval with average 2.56 calls per turn
- Pruning decisions: Self-edited context indicating which documents to remove to maintain retrieval quality
Capabilities
The model breaks complex questions into targeted subqueries rather than attempting to answer everything at once. It can make multiple tool calls in parallel, averaging 2.56 calls per turn, which reduces the total number of interaction rounds needed. A distinctive capability involves self-editing context by pruning irrelevant documents mid-search with 0.94 accuracy, allowing sustained retrieval quality even within bounded context windows. The model generalizes across different domains, having been trained on web, legal, and finance tasks while performing well on held-out domains and public benchmarks including BrowseComp-Plus, SealQA, FRAMES, and HLE.
What can I use it for?
context-1 works within agentic systems that need reliable document retrieval. Use it for research assistants that need to find supporting evidence across large document collections, legal discovery systems that search contracts and regulations, financial analysis tools that retrieve relevant market data and reports, or customer support systems that pull relevant documentation to answer questions. The speed and cost efficiency make it practical for production systems where retrieval happens frequently. Companies can integrate it as the retrieval backbone of larger AI applications, reducing infrastructure costs compared to using larger models for retrieval tasks alone.
Things to try
Since the model decomposes complex queries into subqueries, test it with questions that have multiple constraints or require information from different document types—the model's strength lies in handling these multi-hop scenarios. Experiment with how the self-pruning mechanism maintains quality across extended search sessions by monitoring which documents it removes and why. You can also evaluate its cross-domain generalization by testing on document collections from domains different from web, legal, and finance to see how well the learned search strategies transfer.
This is a simplified guide to an AI model called context-1 maintained by chromadb. If you like these kinds of analysis, join AIModels.fyi or follow us on Twitter.