NervaPack Performance Benchmarks¶
Last Updated: 2026-07-06 Version Tested: 0.4.1
Executive Summary¶
NervaPack achieves significantly lower token counts while maintaining high retrieval recall on real-world coding issues from the SWE-bench Lite dataset.
| Metric | NervaPack | Aider (Repo Map) | Repomix (Full Pack) |
|---|---|---|---|
| Average Tokens | 2300 | 10300 | 88000 |
| Average Recall@k | 96% | 94% | 100% |
| Token Reduction (vs Repomix) | 97.4% | 88.3% | Baseline |
Token counts refer to the context size provided to the LLM. Recall@k refers to whether the ground-truth modified files required to solve the issue were successfully included in the context.
Head-to-Head Comparison¶
We evaluated NervaPack against standard context generation tools on a sample of 5 public repositories from the SWE-bench Lite dataset.
- NervaPack: Graph-based 2-hop retrieval.
- Aider: Default AST-based repo map generated via dry-run.
- Repomix: Bundles the entire repository contents.
NervaPack achieves 4.5x fewer tokens than an Aider repo map, and 38.3x fewer tokens than a full Repomix bundle, while maintaining competitive recall (missing ground-truth files less than 5% of the time on average).
For full methodology and raw data, see our Methodology page.