SAG2 builds RAG index 7000x faster with no embeddings or GPU, beats HippoRAG 2
udmrzn · x · 2026-11-11
Open-source SAG2 / SAG Wiki (Zleap-AI, Rust, MIT, 819 GitHub stars) claims to upend traditional RAG: indexing requires no LLM calls, no embeddings, no vector store, no GPU — just a BM25 keyword index over raw text.
- Speed: on 11,656 text chunks, the previous generation took 4.57 hours to index; SAG2 takes 2.36 seconds — roughly 7000x faster.
- Semantic retrieval via JEV: BM25 does coarse filtering, JEV judges useful evidence, then picks new keywords from that evidence to keep searching — threads are linked via keywords and passages, no entity extraction needed.
- Results: MuSiQue multi-hop Recall@5 hits 80.68%, 15.55 points above HippoRAG 2; on the 510k-document EnterpriseRAG-Bench, graph-based rivals like LightRAG and HippoRAG 2 failed to finish, while SAG2 indexed in 8m32s with Recall@10 of 81.10%.
- Tool: SAG Wiki ships as 13MB (Windows) / 18MB (macOS) apps — point it at a local folder and it becomes a fully automatic knowledge base.
More from coding & agent
- Mintlify open-sources agent-traffic: agents made 489M doc requests, 77% of all traffic — handotdev · 2026-11-11
- Baseten details deep research agent training: breadth-rewarding RL plus SFT for report writing — baseten · 2026-11-11
- Practical Data: Chrome extension applies AI semantic formatting to web tables — calabi_and_yau · 2026-11-11
- Google Antigravity ships 81 weekly updates: Gemini 4 Carbon and /swarm multi-agent land — rseroter · 2026-11-11
- Agent workflow upgrade: narrated demo videos of each change instead of HTML reports — randal_olson · 2026-11-11
- Voice coffee-ordering demo updates UI in ~200ms mid-sentence — TheMoonMidas · 2026-11-11