RAGFlow: BM25+grep alone hits 91% FRAMES on MiniMax M3, no vector index
dair_ai · x · 2026-11-02
InfiniFlow's RAGFlow team argues retrieval is now a disciplined loop owned by a search harness, not an embedding layer. Their harness with only BM25 and grep answers 85% of BrowseComp-Plus and 91% of multi-hop FRAMES on MiniMax M3 — no vector or compiled index needed.
Key findings:
- Adding summary trees, PageIndex-style outlines, wikis, or graphs on top did not improve FRAMES accuracy; outlines degraded at scale.
- Structures pay off in token cost: a navigation tree whose leaves are claim-evidence pairs cut query tokens by 91% when evidence sufficed (+1.9% when it didn't), averaging a 38% reduction.
- Savings cost retrieval rounds: on one paired subset, tokens halved but latency roughly tripled; a decision gate cut latency from 140s to 30s.
Takeaway: if accuracy is saturated, measure index structures on tokens and latency instead.
Related event: Simple Retrieval Tools Dramatically Boost Agent Accuracy, Papers Show(2 posts)→
More from coding & agent
- Podcast: running 50+ agents, cutting Claude costs 97%, surviving a $550k monthly API bill — Ror_Fly · 2026-11-03
- LLM4Code 2027 workshop at ICSE Dublin opens CFP with Nov 13 deadline — LingmingZhang · 2026-11-03
- Claude Code startup context cut from 63K to 22K tokens, saving 65% — dSebastien · 2026-11-03
- 'Addicted to clicking these buttons': coding agent permission dialogs become an AI-coder meme — ThePeterMick · 2026-11-02
- Users push back on OpenAI merging ChatGPT and Codex usage limits — TinfoilTricorn · 2026-11-02
- Can managing a team of AI agents make you a better engineering manager? — HankYeomans · 2026-11-02