sgrep cuts agent search cost 21% on SWE-bench, and the accuracy gain is within noise
Context.dev's open-source sgrep, a local semantic layer on top of ripgrep, cut LLM API spend by 21.3% in a 100-attempt SWE-bench Verified test while solving 70 attempts to ripgrep's 68. The authors' own confidence interval on that accuracy gap runs from minus 3 to plus 8 points.
9 min read