SAG-Wiki: a local knowledge base that indexes without any LLM calls

Yesterday we covered LEANN — the RAG index that shrank 201GB to 6GB. Today’s entry goes one step further: SAG-Wiki (Zleap-AI, MIT, Rust + Tauri + Svelte 5, created October 18, 2026, 103 stars as of November 6) drops LLM calls and embeddings from indexing entirely. It is a 14MB desktop app (macOS/Windows) that turns local folders into a searchable knowledge base — about 200ms to index a million characters, retrieval via its own SAG2 evidence scoring, answers traceable to source passages, plus a ready Agent Skill and CLI for coding agents. A Chinese write-up framed it bluntly: “RAG is dead this time.”

Core features

• Zero-LLM indexing: no model calls, no embeddings; local file watching keeps the index current.
• SAG2 evidence retrieval: an original retrieval method with an arXiv paper (2610.22588) and an open benchmark repo (SAG-Benchmark); scores supporting passages across files. Semantic-level retrieval needs a configured Jev service.
• Cited Q&A: configure an answer model to ask across files; answers cite source locations you can open and verify.
• Desktop toolkit: Markdown reading/editing, HTML preview, related-document exploration, global-shortcut quick search, terminal search; PDF/Office/EPUB convert out of the box, with image recognition and AV transcription as optional components.
• Agent-ready: CLI plus a ready SKILL.md, so a coding agent shares the same knowledge base.

Typical use cases

• A personal document hub: notes, docs and code in one searchable folder with traceable citations.
• Memory for coding agents: the project knowledge base is shared with the agent via the skill, and retrieval burns no LLM quota.
• Privacy-first search: basic indexing and search need no API key and never leave the machine, and the macOS build is unsigned so expect a first-run warning.

Quick start

Grab the desktop build from GitHub Releases (macOS/Windows, 14MB, unsigned so macOS shows a first-run prompt) and drop a folder in to index. For agent use, install the SKILL.md. Semantic SAG2 retrieval and cited Q&A need a Jev service and API key — plain keyword search is entirely free and local.

Summary

SAG-Wiki fits individuals and agent-heavy users who want instant indexing, traceable citations and a zero-cost baseline; deep semantic Q&A requires the Jev service, and versus the LLM-Wiki pattern (Karpathy’s proposal) it trades auto-generated interlinked pages for direct original-document search. Facts and limits: MIT, Rust/Tauri, created October 18, 2026 with just 103 stars — early days; the “beats RAG/GraphRAG” claim is from the project’s own benchmark, pending independent validation. It answers the same question as yesterday’s LEANN from the opposite direction: LEANN stores nothing and computes at query time, SAG-Wiki skips vectors altogether — both point at the same verdict that the bottleneck of local RAG is not the model but the index engineering.