Knowledge-as-Skill：LLM 智能体自主使用知识库的结构设计
原标题：Knowledge-as-Skill: A Structural Design for Autonomous Knowledge-Base Use by LLM Agents
AI 摘要
该论文提出 Knowledge-as-Skill 结构设计，让 LLM 智能体能够自主发现、导航和使用知识库，而非依赖传统 RAG 的固定检索-拼接-生成流程。设计包含 SKILL.md 发现层、每目录 index.md 导航层和带 YAML frontmatter 的知识层，遵循 Open Knowledge Format 与 Skill 协议。在 WixQA 企业客服基准上初步评测得到 0.889 Factuality 和 0.816 Context Recall，高于同期 Corpus2Skill 的 0.767 和 0.708，但 Faithfulness 和 Context Precision 略低、交互轮次更多，作者强调这是方向性证据而非因果比较。
正文节选
Knowledge-as-Skill: A Structural Design for Autonomous Knowledge-Base Use by LLM Agents Abstract Retrieval-augmented generation (RAG) is the dominant approach for giving large language models (LLMs) access to external knowledge. Its conventional “retrieve–concatenate–generate” pipeline, however, makes the retrieval decision on behalf of the model: evidence is retrieved and injected whether or not the question requires it. As tool use and agent loops become more reliable, decisions about whether