Ask HN: 语境构建方法的状态 (SoTA)
2 分•作者: h4ch1•4 个月前
Hi HN，我是一个极简 AI 用户（我使用带有标准 vscode 聊天功能，没有 MCPs 来构思计划），但最近我一直在接触构建上下文的 MCPs/工具，这些工具可以构建你的代码库的知识图谱。
这真的很有趣，因为我同时在处理相当大的代码库，而且我的助手不必一遍又一遍地重新索引和重新阅读所有文件，这非常吸引人。
我的问题是，你们使用什么方法来索引并将这些方法连接到你们的助手？
我的研究让我发现了以下内容，但我基本上对衡量它们有多好感到非常困惑。
1. [https://github.com/abhigyanpatwari/GitNexus](https://github.com/abhigyanpatwari/GitNexus) - 似乎是最受欢迎的
2. [https://github.com/DeusData/codebase-memory-mcp](https://github.com/DeusData/codebase-memory-mcp) - 也很吸引人/新颖
3. [https://github.com/JaredStewart/coderlm](https://github.com/JaredStewart/coderlm) - 基于 tree-sitter 的方法看起来非常好。
很想知道积极使用前沿模型和方法的人是如何在这个领域中进行探索的。
查看原文
Hi HN, I am a very minimal AI user (I use stock vscode chat with no MCPs to ideate on plans), but recently I've been coming across context building MCPs/tools that build a knowledge graph of your codebase.<p>This is really interesting because I am working on pretty large codebases simultaneously and the idea that my agent won't have to re-index and re-read all my files over and over again is pretty enticing.<p>My question is what methods are you using to index and connect these to your agents?<p>My research has led me to the following but I am basically very confused by what metric to measure how good they are.<p>1. https://github.com/abhigyanpatwari/GitNexus - seems like the most popular<p>2. https://github.com/DeusData/codebase-memory-mcp - also interesting/new<p>3. https://github.com/JaredStewart/coderlm - the tree-sitter based approach seems really good.<p>Would love to know how people actively using frontier models and methods navigate this domain.