A wiki that indexes a million characters in 200 ms, no LLM required

What it does

SAG-Wiki is a lightweight desktop app (Tauri + Svelte, Rust core) that turns local folders into a searchable knowledge base. You add a folder, it indexes your documents, notes, and code, and you can search passages, ask cited questions, and jump to the exact source behind each answer. A CLI and an installable Skill expose the same index to agents like Claude Code, Codex, and OpenCode.

The interesting bit

Indexing and retrieval use no embeddings and no LLM calls — the whole pipeline runs locally against a SQLite index, with a claimed 14 MB download and roughly 200 ms to index a million characters. The retrieval method, SAG2, is published as an arXiv paper with benchmark code, and the README reports Recall@10 beating SAG and BM25 on MuSiQue, 2WikiMultiHopQA, and HotpotQA.

Key highlights

• Runs fully offline for indexing and quick search; no API key needed for basic use
• Answers come with citations that highlight the original passage in the source file
• Watches local files and updates the index as they change
• PDF, Office, and EPUB convert out of the box; image recognition and transcription are optional downloads
• The Rust CLI is the runtime foundation — desktop and terminal share one service and one index

Caveats

• Semantic retrieval and cited Q&A are not free-standing: they require configuring a “Jev” service (with its own state/questions/answers.*.noul protocol) plus an answer model, so the no-LLM claim applies to indexing, not the full Q&A experience
• In-depth retrieval and Q&A send questions and passages to whatever services you configure
• The npm CLI ships binaries only for macOS Apple Silicon and Windows x64

Verdict

Worth a look if you want a fast, local, source-traceable knowledge base shared between you and your coding agent. If you expected fully offline semantic Q&A out of the box, temper that — the clever indexing is local, the clever answering is not.