SAG-Wiki

English · 简体中文 · Changelog

Lightweight and easy. Fast indexing. Precise retrieval.

A lightweight knowledge base for people and agents. Index and search your local files without LLM calls, then ask questions with answers you can trace to the source.

https://github.com/user-attachments/assets/e9c5499d-6057-4add-ac7e-f57eae381a2f

PDF basic text can become searchable before background OCR and table enhancement finish. Independent basic indexing keeps supported local Office/RTF documents moving, while PDF segments reuse loaded models. See the changelog for release-specific changes and the bilingual documentation for usage. The download link below points to the latest published stable Release.

Contents

Project · Technology · User Guide · Agent Guide · Developer Guide · Community

SAG-Wiki in one minute

Download the latest desktop release for macOS and Windows

Install for your agent

Install https://github.com/Zleap-AI/SAG-Wiki/blob/main/cli/skills/sag/SKILL.md and follow its setup flow for my current project.

The installer size excludes optional document conversion components and recognition models downloaded on demand.

Add a folder to make your documents, notes, and code searchable. Find a passage, ask a question across files, open the source behind an answer, or give your coding agent access to the same knowledge base.

• Light: no separate database to install; optional recognition components download separately.
• Fast: index a million characters in approximately 200 ms, with no LLM calls for indexing or retrieval.
• Precise: our original SAG2 retrieval technology outperforms RAG, GraphRAG, and other approaches.

The indexing figure above concerns text indexing, excluding document conversion, OCR and table recognition; it is not a whole-PDF completion-time estimate.

Read and edit Markdown, preview HTML, explore related documents, or search from the terminal. PDF, Office, and EPUB documents convert out of the box; image recognition and audio/video transcription are optional components.

For image-heavy DOCX/PPTX files where only text is needed, enable Office text only in Settings → File settings (off by default). Images and embedded objects are skipped before loading; text inside images is not recognized. See the folder guide.

How it compares

The comparison uses Karpathy's LLM Wiki pattern and the standard vector RAG workflow. Features vary by implementation.

Technology

SAG-Wiki runs on SAG2, our original retrieval method. It builds a local index without embeddings or LLM calls, then uses Jev evidence scoring and query-time expansion to find supporting passages across documents. LLMs are used when you choose to generate an answer.

https://github.com/user-attachments/assets/bcdf9dbf-7535-41d6-bd22-bde0ea824525

_{SAG-Wiki is represented by its underlying SAG2 benchmark results. Recall@10 measures evidence recall in the first ten results. SAG2 uses 1,000 questions per benchmark; MuSiQue was used for method development. SAG and BM25 are reference values from the original SAG paper. Difficulty labels are presentation categories; relative gains use the displayed scores.}

Read the SAG2 paper · For the method, benchmark results, and reproduction code, see SAG-Benchmark.

User Guide

SAG-Wiki itself is free and open source. External service charges depend on your chosen provider. Original files stay on your machine; in-depth retrieval and Q&A send the required questions, document passages, or context to the services you configure.

1. Download and add a folder

Download the macOS DMG from the latest release link above, open it, and drag SAG Wiki into Applications. The macOS app, bundled CLI, and DMG are Developer ID signed and notarized by Apple.

Install and open SAG-Wiki, switch the sidebar to Folders, then click + beside folder management to add the folder you want to search. Your files stay in their original locations.

2. Search your files

Use Search as soon as basic text is indexed. For mixed or scanned PDFs, OCR and table enhancement continue in the background; file-list states show whether basic text is searchable, enhancement is pending/running, or enhancement failed. Existing text remains searchable after an enhancement failure; image-only scans become searchable only after recognized text is indexed. Wait for completion and check the original for full coverage. Quick search works locally; configure a decision model in Settings → Models for in-depth semantic retrieval.

For a local Jev-compatible decision model, select Custom and enter its model name and base URL (for example, http://127.0.0.1:8000/v1), or its complete /systemone or /decisions endpoint. Choose No authentication under Decision model authentication if the service does not require an API key; this omits the authorization header and removes the saved decision model key. The service must implement Jev’s state, questions, and answers.*.noul protocol. For the CLI, run sag config --url http://127.0.0.1:8000/v1 --model YOUR_MODEL --auth none. To switch back to a keyed service, use --auth bearer --key-stdin and enter that service’s key.

3. Ask questions and check the source

Configure your answer model in Settings → Models, then ask in Chat. Use @ to select files, and click a citation to see the passage supporting the answer.

Agent Guide

Give your agent this one sentence:

Install https://github.com/Zleap-AI/SAG-Wiki/blob/main/cli/skills/sag/SKILL.md and follow its setup flow for my current project.

SAG supports Claude Code, Codex, and OpenCode. The Skill includes CLI installation, configuration, project indexing and verification. The CLI works without SAG Wiki desktop and shares its settings and knowledge base when the desktop is installed. Existing settings are reused; API-key-authenticated semantic search needs your own decision model key, entered privately in a terminal. The GitHub Skill link is the canonical installation entry; the agent retrieves its supporting files from the same directory.

With Node.js 18 or later installed:

npm install -g @zleap-ai/sag

The published npm package includes macOS Apple Silicon, Windows x64 and Linux ARM64 / x86_64 binaries (Linux verified on Ubuntu 24.04 and Debian 12). It requires no Rust toolchain or source compilation. Installing the CLI does not install the desktop app.

sag --version
sag

sag add
sag check
sag "Where is permission checked?" . -m 5

Developer Guide

The Rust CLI is the runtime foundation. The sag executable starts a shared local service; both terminal commands and the Tauri desktop connect to it through the same client. Indexing, retrieval, and model configuration live in one core, and remain available after the desktop exits.

flowchart TD
    Desktop["Desktop · Svelte + Tauri"] --> Client["sag-client"]
    Agent["Terminal / Agent · sag CLI"] --> Client
    Client --> Service["sag service serve · Rust core"]
    Service --> Local["Local files + SQLite index"]
    Service --> Jev["Jev · evidence retrieval"]
    Service --> LLM["LLM · answer generation"]

Follow the environment guide to prepare Node.js, Rust, and your platform's Tauri development dependencies, then run:

git clone https://github.com/Zleap-AI/SAG-Wiki.git
cd SAG-Wiki
npm ci
npm run tauri dev

After changes, run npm run check and npm test; the checks section lists all validation commands. See the development guide, architecture, and documentation index for builds and implementation details.

Contributing and License

Contributions are welcome through Issues and Pull Requests. For changes to the application, follow the development guide; retrieval research and reproduction code live in SAG-Benchmark.

SAG-Wiki is released under the MIT License. Third-party notices are in src-tauri/licenses.

Community

Join the SAG community on Discord or WeChat to connect with maintainers and other users.

Discord	WeChat