上下文即基础设施：智能体系统检索中的知识编译与Harness
原标题：Context as Infrastructure: Knowledge Compilation for Retrieval in Agentic Systems with Harnesses
AI 摘要
该论文提出，在检索增强生成（RAG）转向智能体化后，检索准确率问题已基本由搜索 harness 解决（仅用 BM25 和 grep 配合 MiniMax M3 即在 BrowseComp-Plus 达 85%、FRAMES 达 91%），但暴露出上下文浪费问题。作者主张缺失的一层是「知识编译」：用 LLM 在入库阶段为文档补充导航树、实体图、时间线、本体等结构化产物。实验显示，以「主张-证据对」为叶子的导航树在证据充分时减少 91% 查询 token，平均降低 38%，决策门将延迟从 140 秒降至 30 秒。作者还描述了 RAGFlow 的编译流水线与检索 harness，二者共同构成「智能体原生数据库」。
正文节选
Context as Infrastructure: Knowledge Compilation for Retrieval in Agentic Systems with Harnesses Abstract. Retrieval-augmented generation is becoming agentic. A search harness writes queries, reads what it finds, reformulates, and decides when to stop, and this shift has largely solved the accuracy problem: a harness whose only retrieval tools are BM25 and grep, with MiniMax M3 as the answering model, answers 85% of BrowseComp-Plus and reaches 91% on the multi-hop FRAMES benchmark—two benchmarks