Chroma Context-1: Training a Self-Editing Search Agent
12 hours ago
- #agentic-search
- #retrieval-augmented-generation
- #multi-hop-retrieval
- Introduction of Context-1, a 20B parameter agentic search model for retrieval-augmented-generation (RAG).
- Context-1 addresses limitations of single-stage retrieval by enabling multi-hop retrieval through iterative query decomposition and evidence gathering.
- The model is trained on synthetic tasks, focusing on planning and evaluation skills, with a curriculum shifting from recall to precision.
- Context-1 features self-editing context management, allowing selective retention or discarding of retrieved documents to manage context window size.
- Performance benchmarks show Context-1 matches or exceeds frontier models in retrieval tasks across web, finance, legal, and email domains.
- The model is released as open weights, along with the synthetic task generation pipeline, to support reproducibility and further research.
- Future directions include expanding task diversity, improving tool use and search infrastructure, and enhancing context management strategies.