Alibaba's open-code-review: The Same Claude Model, 4.7x Better Precision Than Claude Code

说明与用途

阿里巴巴的 open-code-review（OCR）将确定性规则与 LLM 结合，使用同一 Claude 4.6 Opus 模型实现 33.9% 的精准率，比 Claude Code 高 4.7 倍，耗时和 token 数量均降低 10 倍。

问题

Developers need a reliable, high‑precision AI code review tool that can be run locally or in CI pipelines without relying on a single vendor’s model.

结果

OCR achieves 4.7× higher precision than Claude Code while using 10× fewer tokens and 10× less time, demonstrating that a deterministic pre‑filter harness can dramatically improve LLM‑based code review quality.

工作流

• Install OCR globally: npm i -g @alibaba-group/open-code-review
• Configure an endpoint (Anthropic, OpenAI, or local) via OCR config
• Run OCR against a diff or repository: ocr <repo-or-diff>
• Review line‑level comments produced by the LLM
• Iterate: adjust deterministic rules or switch model endpoint as needed

证据

OCR posts 33.9% precision versus Claude Code's 7.23%, in 1m 23s versus 13m 6s, using 385K tokens versus 5.66M.

来源

The hybrid design pairs deterministic language rulesets with an LLM agent that adds precise line‑level explanations.

来源