Chroma Context-1: Training a Self-Editing Search Agent
Summary
Chroma Context-1 introduces a 20B-parameter agentic search model designed to operate as a retrieval subagent that iteratively decomposes queries, searches across multiple turns, and self-edits its context to manage a bounded token budget. It demonstrates near frontier-level retrieval performance at lower cost and faster inference, leveraging a training pipeline that includes synthetic data generation, supervised fine-tuning, and reinforcement learning with CISPO. The paper details the agent harness, context-management mechanisms, benchmarks, and future directions for this approach to scalable, cost-efficient AI-powered search.