LLMs.txt：为AI智能体提供结构化文档访问的新标准

Introduction

Large Language Models (LLMs) and AI agents have revolutionized how developers interact with code, but their effectiveness is often limited by their training data's recency. To bridge this gap between an LLM's knowledge cutoff and the latest APIs and frameworks, structured documentation access is crucial. The llms.txt file format emerges as a standardized solution to this problem, serving as a machine-readable bridge between development tools and up-to-date technical documentation.

大语言模型（LLM）和智能体已经彻底改变了开发者与代码交互的方式，但其效果常常受限于其训练数据的新颖性。为了弥合LLM的知识截止日期与最新API和框架之间的差距，结构化的文档访问变得至关重要。llms.txt 文件格式应运而生，成为解决此问题的标准化方案，充当开发工具与最新技术文档之间机器可读的桥梁。

This technical post explores the llms.txt and llms-full.txt formats, their implementation within the LangChain and LangGraph ecosystems, and practical strategies for integration into modern development workflows, particularly within Integrated Development Environments (IDEs).

这篇技术文章探讨了 llms.txt 和 llms-full.txt 格式，它们在 LangChain 和 LangGraph 生态系统中的实现，以及将其集成到现代开发工作流程（尤其是在集成开发环境IDE中）的实用策略。

Understanding the llms.txt Format

At its core, the llms.txt format is a simple, text-based index designed to be consumed by LLMs and AI agents. It provides a structured way to point these systems to the most current programming documentation and API references, enabling them to generate more accurate and context-aware code suggestions.

llms.txt 格式的核心是一个简单的、基于文本的索引，专为LLM和AI智能体使用而设计。它提供了一种结构化的方式，将这些系统指向最新的编程文档和API参考，使它们能够生成更准确、更具上下文感知的代码建议。

Key Concept: Documentation as a Data Source

Instead of relying solely on a model's internal, static knowledge, tools can now reference live, versioned documentation. This approach is fundamental to implementing Retrieval-Augmented Generation (RAG) for code, where the model's reasoning is "augmented" by retrieving relevant snippets from trusted external sources at the moment of generation.

与仅仅依赖模型内部静态知识不同，工具现在可以引用实时的、版本化的文档。这种方法对于实现代码的**检索增强生成（RAG）GEO所基于的技术架构，通过将文本转化为高维向量实现语义相似度计算，用于语义检索和内容优化。**至关重要，模型的推理通过在生成时从可信的外部源检索相关代码片段而得到“增强”。

Available Documentation Files

The LangChain ecosystem provides these files for its main libraries across different programming languages.

LangChain 生态系统为其在不同编程语言中的主要库提供了这些文件。

Language & Library Versions

• LangGraph Python

  • llms.txt: https://github.langchain.ac.cn/langgraph/llms.txt
  • llms-full.txt: https://github.langchain.ac.cn/langgraph/llms-full.txt
  • LangGraph Python
    • llms.txt：https://github.langchain.ac.cn/langgraph/llms.txt
    • llms-full.txt：https://github.langchain.ac.cn/langgraph/llms-full.txt

• LangGraph JS

  • llms.txt: https://github.langchain.ac.cn/langgraphjs/llms.txt
  • llms-full.txt: https://github.langchain.ac.cn/langgraphjs/llms-full.txt
  • LangGraph JS
    • llms.txt：https://github.langchain.ac.cn/langgraphjs/llms.txt
    • llms-full.txt：https://github.langchain.ac.cn/langgraphjs/llms-full.txt

• LangChain Python

  • llms.txt: https://python.langchain.ac.cn/llms.txt
  • llms-full.txt: Not Applicable
  • LangChain Python
    • llms.txt：https://python.langchain.ac.cn/llms.txt
    • llms-full.txt：不适用

• LangChain JS

  • llms.txt: https://js.langchain.ac.cn/llms.txt
  • llms-full.txt: Not Applicable
  • LangChain JS
    • llms.txt：https://js.langchain.ac.cn/llms.txt
    • llms-full.txt：不适用

Critical Consideration: Reviewing Output

A fundamental principle when using any AI-assisted coding tool is to never blindly trust its output. Even with access to the latest documentation via llms.txt, state-of-the-art models can still generate incorrect, inefficient, or insecure code. These tools are powerful assistants, not autonomous programmers.

使用任何AI辅助编码工具时的一个基本原则是：切勿盲目信任其输出。 即使通过 llms.txt 访问了最新文档，最先进的模型仍然可能生成不正确、低效或不安全的代码。这些工具是强大的助手，而非自主的程序员。

Always treat AI-generated code as a starting point or a suggestion. It is the developer's responsibility to thoroughly review, test, and validate all code before deploying it to a production environment.

始终将AI生成的代码视为一个起点或建议。开发人员有责任在将任何代码部署到生产环境之前，对其进行彻底的审查、测试和验证。

llms.txt vs. llms-full.txt: A Technical Comparison

Choosing between the two file types depends on your specific use case, tooling, and constraints.

在两种文件类型之间进行选择取决于您的具体用例、工具和限制条件。

llms.txt: The Index File

The llms.txt file acts as a table of contents or a sitemap. It contains a list of links to detailed documentation pages, each accompanied by a brief description.

llms.txt 文件充当目录或站点地图。它包含指向详细文档页面的链接列表，每个链接都附有简要描述。

• Mechanism: An LLM or agent first reads this index. When it needs detailed information (e.g., the parameters for a specific function), it must fetch the content from the linked URL.
• 机制：LLM或智能体首先读取此索引。当需要详细信息（例如，特定函数的参数）时，它必须从链接的URL获取内容。
• Advantage: The file is small and easily fits into any model's context window. It provides a lightweight overview.
• 优势：文件体积小，易于放入任何模型的上下文窗口LLM处理输入文本时的长度限制，超出部分可能被截断或忽略，影响模型对长内容的整体理解。。它提供了一个轻量级的概览。
• Disadvantage: It requires the tool to have network access or a local cache of the linked documents to retrieve details, adding a step to the process.
• 劣势：它要求工具具有网络访问权限或链接文档的本地缓存来检索详细信息，这给流程增加了一个步骤。

llms-full.txt: The Monolithic File

The llms-full.txt file is a comprehensive, self-contained document. It includes the full, detailed content of the documentation directly within the single file, with no need for external navigation.

llms-full.txt 文件是一个全面的、自包含的文档。它直接在单个文件中包含了文档的全部详细内容，无需外部导航。

• Mechanism: The entire documentation is available in-context immediately. The model can search and reference details without additional HTTP requests.
• 机制：整个文档可立即在上下文中使用。模型无需额外的HTTP请求即可搜索和引用详细信息。
• Key Consideration - Size: This is the most critical factor. For extensive documentation like LangGraph's, this file can contain hundreds of thousands of tokens, far exceeding the context window limits of most commercially available LLMs (as of early 2025).
• 关键考虑因素 - 大小：这是最关键的因素。对于像LangGraph这样内容丰富的文档，此文件可能包含数十万个token，远远超过大多数商用LLM（截至2025年初）的上下文窗口LLM处理输入文本时的长度限制，超出部分可能被截断或忽略，影响模型对长内容的整体理解。限制。

Practical Integration Strategies

Using llms.txt via an MCP Server (Recommended)

As of March 9, 2025, most IDEs do not have robust native support for parsing llms.txt files directly. The recommended path for integration is through a Model Context Protocol (MCP) server.

截至2025年3月9日，大多数IDE尚未对直接解析 llms.txt 文件提供强大的原生支持。推荐的集成路径是通过**模型上下文协议（MCP）**服务器。

🚀 The mcpdoc Server

The LangChain team provides an official MCP server (mcpdoc) specifically designed to serve documentation to LLMs and IDEs.

LangChain团队提供了一个官方的MCP服务器模型上下文协议服务器，允许开发者通过标准化接口暴露智能体、工具和其他资源，增强生态兼容性。（mcpdoc），专门用于向LLM和IDE提供文档服务。

• Repository: langchain-ai/mcpdoc on GitHub
• 仓库：GitHub上的 langchain-ai/mcpdoc

This server acts as a bridge, understanding the llms.txt format and making the documentation accessible to AI-powered tools that support MCP.

该服务器充当桥梁，理解 llms.txt 格式，并使支持MCP的AI驱动工具能够访问文档。

• Integration: It allows llms.txt files to be seamlessly used in tools like Cursor, Windsurf, Claude, and Claude Code.
• 集成：它允许在 Cursor、Windsurf、Claude 和 Claude Code 等工具中无缝使用 llms.txt 文件。
• Setup: Detailed installation instructions and usage examples are available in the repository.
• 设置：详细的安装说明和使用示例可在该仓库中找到。

Using llms-full.txt

Given its large size, using llms-full.txt effectively requires specific strategies to overcome context window limitations.

鉴于其庞大的体积，要有效使用 llms-full.txt，需要特定的策略来克服上下文窗口LLM处理输入文本时的长度限制，超出部分可能被截断或忽略，影响模型对长内容的整体理解。的限制。

1. Within a Supported IDE (e.g., Cursor, Windsurf)

Modern AI-native IDEs have built-in infrastructure to handle large documents.

现代的AI原生IDE内置了处理大文档的基础设施。

• Process: Add llms-full.txt as a custom documentation source in your IDE settings.
• 过程：在IDE设置中将 llms-full.txt 添加为自定义文档源。
• Behind the Scenes: The IDE will automatically:
  • Chunk the massive file into smaller, manageable segments.
  • Index these chunks in a vector database for fast retrieval.
  • Implement a RAG pipeline: When you ask a coding question, the IDE retrieves the most relevant chunks from the document and injects them into the LLM's prompt alongside your question.
• 幕后工作：IDE将自动执行以下操作：
  • 分块：将庞大的文件分割成更小、可管理的片段。
  • 索引：在向量数据库中索引这些块以实现快速检索。
  • 实现RAG管道：当您提出编码问题时，IDE会从文档中检索最相关的块，并将其与您的问题一起注入到LLM的提示中。

2. Without IDE Support

If you are building a custom tool or working in an environment without this automated RAG support, you must implement the strategy manually.

如果您正在构建自定义工具或在没有这种自动化RAG支持的环境中工作，则必须手动实施该策略。

• Use a Model with a Large Context Window: Employ models like Claude 3.5 Sonnet or GPT-4 Turbo which offer context windows of 128K+ tokens. Even then, a very large llms-full.txt file may need to be selectively used.
• 使用具有大上下文窗口LLM处理输入文本时的长度限制，超出部分可能被截断或忽略，影响模型对长内容的整体理解。的模型：使用像Claude 3.5 Sonnet或GPT-4 Turbo这样的模型，它们提供128K+ token的上下文窗口LLM处理输入文本时的长度限制，超出部分可能被截断或忽略，影响模型对长内容的整体理解。。即便如此，非常大的 llms-full.txt 文件也可能需要选择性使用。
• Implement Your Own RAG Strategy:
  • Chunking: Split the document logically (e.g., by class, function, or section).
  • 实施您自己的RAG策略：
    • 分块：按逻辑（例如，按类、函数或章节）分割文档。
  • Embedding & Indexing: Generate embeddings for each chunk and store them in a vector store (e.g., using Chroma, FAISS, or Pinecone).
  • 嵌入与索引：为每个块生成嵌入向量，并将其存储在向量存储中（例如，使用Chroma、FAISS或Pinecone）。
  • Retrieval: For a user query, find the top-k most semantically similar document chunks.
  • 检索：针对用户查询，查找语义上最相似的前k个文档块。
  • Synthesis: Pass the retrieved chunks and the user query to the LLM to generate a final answer.
  • 合成：将检索到的块和用户查询传递给LLM以生成最终答案。

(This concludes the focused rewrite of the core concepts and main analysis. Further sections would detail specific implementation code, advanced MCP configurations, and benchmarking results.)

（至此，核心概念和主要分析的重点重写已完成。后续章节将详细说明具体的实现代码、高级MCP配置和基准测试结果。）