Chroma Context-1: Training a Self-Editing Search Agent

gist

Chroma's Context-1 is a 20B gpt-oss search subagent that returns ranked supporting documents instead of answers. Trained on more than eight thousand synthetic multi-hop tasks, it decomposes queries, searches iteratively, and prunes its own context so long-horizon retrieval stays inside a bounded window. The authors report frontier-comparable retrieval at a fraction of the cost and up to 10x faster inference, then release the weights and data-generation pipeline.

ideas

• Separate search from generation. Context-1 returns a ranked document set to a downstream reasoner. Missing evidence is worse than extra noise, because the generator can filter but cannot invent a document it never received.
• Teach the agent to edit its own context. A prune tool plus token-budget pressure lets the model discard irrelevant chunks mid-search instead of summarizing them, reducing context rot without lossy compression.
• Train recall first, then precision. Supervised warmup on large-model traces is followed by RL with a recall-heavy F1 that later anneals toward precision, plus trajectory-recall credit for documents later pruned.
• Verify synthetic tasks by extraction. The judge extracts document and clue quote pairs and checks they appear verbatim, reaching more than 80% human alignment without full-document labeling.
• Small specialized subagents generalize. Trained only on web, legal, and finance, Context-1 still improves on held-out email and public benchmarks. Four parallel rollouts stay cheaper than one frontier call.

quotes