Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,73 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
base_model:
|
| 4 |
+
- openai/gpt-oss-20b
|
| 5 |
+
---
|
| 6 |
+
|
| 7 |
+
# Chroma Context-1
|
| 8 |
+
|
| 9 |
+
Context-1 is a 20B parameter agentic search model trained
|
| 10 |
+
to retrieve supporting documents for complex, multi-hop
|
| 11 |
+
queries. It is designed to be used as a retrieval subagent
|
| 12 |
+
alongside a frontier reasoning model: given a query,
|
| 13 |
+
Context-1 decomposes it into subqueries, iteratively
|
| 14 |
+
searches a corpus, and selectively edits its own context
|
| 15 |
+
to free capacity for further exploration.
|
| 16 |
+
|
| 17 |
+
Context-1 achieves retrieval performance comparable to
|
| 18 |
+
frontier LLMs at a fraction of the cost and up to 10x
|
| 19 |
+
faster inference speed.
|
| 20 |
+
|
| 21 |
+
**Technical report:**
|
| 22 |
+
[Chroma Context-1: Training a Self-Editing Search Agent](https://trychroma.com/research/context-1)
|
| 23 |
+
|
| 24 |
+
## Model Details
|
| 25 |
+
|
| 26 |
+
- **Base model:** gpt-oss-20b
|
| 27 |
+
- **Parameters:** 20B (Mixture of Experts)
|
| 28 |
+
- **Training:** SFT + RL (CISPO) with a staged curriculum
|
| 29 |
+
- **Precision:** BF16 (MXFP4 quantized checkpoint coming soon)
|
| 30 |
+
|
| 31 |
+
## Key Capabilities
|
| 32 |
+
|
| 33 |
+
- **Query decomposition:** Breaks complex multi-constraint
|
| 34 |
+
questions into targeted subqueries.
|
| 35 |
+
- **Parallel tool calling:** Averages 2.56 tool calls per
|
| 36 |
+
turn, reducing total turns and end-to-end latency.
|
| 37 |
+
- **Self-editing context:** Selectively prunes irrelevant
|
| 38 |
+
documents mid-search to sustain retrieval quality over
|
| 39 |
+
long horizons within a bounded context window (0.94
|
| 40 |
+
prune accuracy).
|
| 41 |
+
- **Cross-domain generalization:** Trained on web, legal,
|
| 42 |
+
and finance tasks; generalizes to held-out domains and
|
| 43 |
+
public benchmarks (BrowseComp-Plus, SealQA, FRAMES,
|
| 44 |
+
HLE).
|
| 45 |
+
|
| 46 |
+
## Important: Agent Harness Required
|
| 47 |
+
|
| 48 |
+
Context-1 is trained to operate within a specific agent
|
| 49 |
+
harness that manages tool execution, token budgets, context
|
| 50 |
+
pruning, and deduplication. **The harness is not yet
|
| 51 |
+
public.** Running the model without it will not reproduce
|
| 52 |
+
the results reported in the technical report.
|
| 53 |
+
|
| 54 |
+
We plan to release the full agent harness and evaluation
|
| 55 |
+
code soon. In the meantime, the technical report describes
|
| 56 |
+
the harness design in detail.
|
| 57 |
+
|
| 58 |
+
## Citation
|
| 59 |
+
|
| 60 |
+
```bibtex
|
| 61 |
+
@techreport{bashir2026context1,
|
| 62 |
+
title = {Chroma Context-1: Training a Self-Editing Search Agent},
|
| 63 |
+
author = {Bashir, Hammad and Hong, Kelly and Jiang, Patrick and Shi, Zhiyi},
|
| 64 |
+
year = {2026},
|
| 65 |
+
month = {March},
|
| 66 |
+
institution = {Chroma},
|
| 67 |
+
url = {https://trychroma.com/research/context-1},
|
| 68 |
+
}
|
| 69 |
+
```
|
| 70 |
+
|
| 71 |
+
## License
|
| 72 |
+
|
| 73 |
+
Apache 2.0
|