Agentic RAG Struggles to Find All Evidence for a Topic
The ai-study-helper-v2 project tested prompts, query expansion, corpus-grounded facets, auditing, reranking, and section routing for exhaustive topic coverage in Agentic RAG. Adding more tools, turns, and queries did not reliably reveal every semantic facet, while a term-by-term full-text search recovered all three regression probes but retrieved 511 of 526 blocks. The project is therefore preserving block retrieval and moving toward precompiled, source-traceable LLM Wiki evidence maps as a possible way to organize and complete topic coverage.
The experiments show that high recall alone does not provide useful retrieval or prove topic completeness, making coverage evaluation and evidence organization important engineering problems for Agentic RAG systems.
30 秒判断：如果你也遇到“检索结果看起来都相关，但关键子主题总是漏掉”，这篇文章的价值是两个：我们替你系统试过了提示词、查询展开、facet、审计、reranker 和章节路由，并标明各自为什么不够；同时用近期论文说明，这不是单靠把项目参数再调一调就能解决的问题。 适合谁读：正在做 Agentic RAG、知识库问答、主题研究、长答案检索或 LLM Wiki 的工程师和产品负责人。 系列位置：本文接续 05 · Agentic RAG 检索侧落地。前文讲块级工具如何形成可靠的 Agent 检索链路；本篇讨论当问题从“找相关证据”变成“围绕主题找全”时，这条链路在哪里失效。 代码库：lipeidonggz/ai-study-helper-v2 研究证据：065 主题覆盖实验 · 066 章节路由实验 · 冻结 tag 证据标记：✅ 项目实测；🧪 可回归；🤔 当前判断。本文引用的公开论文和 WeKnora 实验不是本项目实测结果，会在对应段落单独标明。
Some source content is omitted. Continue reading on the original site.
#ai-study-helper-v2#Agentic RAG#LLM Wiki#WeKnora#GraphRAG#CoveR#Decompose#Q-CARE#TREC CAR