RAGFlow 团队实测：BM25 加 grep 也能撑起智能体检索
RAGFlow 团队用 MiniMax M3 做了个实验：不建向量索引，光靠 BM25 和 grep 就能答对 91% 的 FRAMES。给智能体检索选型的人省了不少冤枉路。
InfiniFlow 的 RAGFlow 团队在 MiniMax M3 上只用 BM25 和 grep 跑智能体检索，BrowseComp-Plus 达 85%、FRAMES 达 91%，未用任何向量或编译索引。随后叠加摘要树、PageIndex 式大纲、wiki 和图结构，FRAMES 准确率没有提升，大纲在规模变大后反而更差。结构的价值在省 token：叶节点为证据对的导航树平均减少 38% 查询 token，证据充分时降幅达 91%。代价是检索轮次增加，一组对照中 token 减半但延迟约增至三倍，加入决策门控后延迟从 140 秒降到 30 秒。
// Context as Infrastructure for Agents //
The RAGFlow team at InfiniFlow ran a harness with only BM25 and grep on MiniMax M3. It answered 85% of BrowseComp-Plus and 91% of FRAMES with no vector or compiled index.
They then tested summary trees, PageIndex-style outlines, wikis, and graphs on top. None improved FRAMES accuracy, and the outline got worse at scale.
The structures help with context cost. A navigation tree whose leaves are claim and evidence pairs cut query tokens by 91% when the evidence was enough, and added 1.9% when it was not. That averages to a 38% reduction.
The savings cost extra retrieval rounds. On one paired subset, tokens halved while latency roughly tripled. A decision gate brought latency from 140 seconds to 30.
If you are building indexes for agentic RAG, measure them on tokens and latency, since accuracy may already be saturated.
https://t.co/xhFUgtaUHm
Chat with Paper: https://t.co/yw3NrZl0Gw
