Chroma Context-1: Agentic Search Model
Source: https://www.trychroma.com/research/context-1 Author: Chroma Date: 2025-2026
Summary
20B parameter model purpose-built for agentic search — returns ranked supporting documents to a downstream answering model (not answers directly). Trained on gpt-oss-20B via SFT + RL with verifiable rewards. Observe-reason-act loop with four tools: semantic/keyword search, regex, document reading, pruning.
Key Claims
- Architecture: observe-reason-act loop, four tools: search, regex, read, prune
- 2.56 parallel tool calls per turn (vs 1.52 baseline) → 5.2 turns completion (vs 6.7)
- 94.1% accuracy identifying irrelevant documents for removal
- Manages bounded context by selectively discarding tangential info mid-search
- Performance: matches frontier-scale LLMs at 10x faster inference and fraction of the cost
- Apache 2.0 open-source
- Benchmarks: web, finance, legal, email domains — matches models orders of magnitude larger
Connection to Other Sources
Interesting counter to “No RAG for Agents” — Context-1 shows that a specialized search agent with RL training can outperform naive RAG.
Concepts
- Coding Agents — specialized search subagent pattern
- RL Infrastructure — RLVR for search task specialization
- Context Engineering — bounded context management via pruning