Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

RAGFlow paper proposes knowledge compilation for agentic retrieval

AI summary

A paper by the RAGFlow team and East China Normal University argues that agentic retrieval depends heavily on the search harness and proposes compiling source documents into typed artifacts that provide better context to agents. The authors report that BM25 plus grep with MiniMax M3 achieved 85% on BrowseComp-Plus, while a separate harness run reached 91% on FRAMES. In their claim–evidence replacement experiments, total query tokens fell by 91% when evidence was adequate, but end-to-end latency rose roughly threefold on a paired subset of seven questions. These findings are author-reported; the headline token and latency figures came from limited samples, and the decision gate’s calibration audit had not been run.

In-depth insight

The paper presents a concrete architecture for reducing token use in agentic RAG through compiled claim–evidence artifacts, but its benchmark and latency findings require caution because some measurements use limited samples and the decision gate remained uncalibrated.

RAG 走到了十字路口

检索增强生成(RAG)在过去两年里几乎成为了大语言模型落地企业知识场景的“标配”。它通过将外部知识库的检索与生成模型相结合,显著提升了生成内容的时效性和准确性,大幅降低了幻觉率。这个逻辑听起来无懈可击:模型不懂的就现查,查到了再回答。

但现实远比逻辑复杂。

当大量企业真正把 RAG 搬进生产环境后,一个令人沮丧的共识正在形成:RAG 很好,但远不够好。 朴素 RAG——那种“把文档切块、向量化、然后算相似度”的标准流水线——在实际场景中频频翻车。分块策略怎么定?Embedding 模型用哪个?相似度阈值设多少?每一个环节都需要大量调参,调完了依然无法保证检索质量。检索到的文本块与答案不相关,干扰项混入上下文后反而扰乱了 LLM 的推理过程。

#RAGFlow#InfiniFlow#East China Normal University#MiniMax M3#FRAMES#BrowseComp-Plus

Original · 35 min readRead the original
Group chat

No comments yet. Start the conversation.