📄 中文摘要
在对三个真实代码库的15个任务进行基准测试中，结构化的MCP符号检索相比于简单的文件读取实现了95%的平均令牌减少，同时在精确度上达到了96%，而块RAG的精确度仅为74%。该基准测试工具是开源的，用户可以在5分钟内重现所有数据。这项研究是对之前一篇文章的延续，强调了AI代理在代码检索中的有效性和效率。
95% 的令牌减少，96% 的精确度
出处: Bringing The Receipts - 95% AI LLM Token Savings
发布: 2026年3月19日
📄 English Summary
Bringing The Receipts - 95% AI LLM Token Savings
The benchmarking of structured MCP symbol retrieval against naive file reading and chunk RAG across 15 tasks on three real repositories demonstrates a remarkable 95% average token reduction and a precision of 96%, compared to only 74% for chunk RAG. The benchmark harness is open-source, allowing users to reproduce all results in under five minutes. This research serves as a follow-up to a previous article, reinforcing the effectiveness and efficiency of AI agents in code retrieval.