Zleap-AI's open source SAG-Wiki indexes local files without embeddings or LLM calls, then feeds precise cited passages to Claude Code and other agents.
-
SAG-Wiki is an open source local knowledge base that indexes files without embeddings or LLM calls.
-
Indexes about one million characters in 200 ms with a 14 MB installer and no separate database.
-
Built on SAG2 retrieval, which beats SAG and BM25 on MuSiQue, 2WikiMultiHopQA, and HotpotQA recall@10.
-
Ships a Skill for Claude Code, Codex, and OpenCode so agents can query files with citations.
-
Installable via
npm install -g @zleap-ai/sagwith prebuilt macOS and Windows binaries. -
MIT licensed, Rust core, with paper andbenchmark repo available. SAG-Wiki is an open source knowledge base that indexes local files without embeddings, a vector database, or model calls during indexing. Developers can point it at a folder, search the resulting index from a desktop app or CLI, and retrieve citations to exact source passages.
The project combines a Rust runtime, a Tauri desktop app, and agent integrations for Claude Code, Codex, and OpenCode. It has attracted more than 1,000 GitHub stars and targets workflows that need frequent, traceable searches across changing files.
Vectors leave the indexing path #
SAG-Wiki’s installer is about 14 MB and requires no separate database. The project reports indexing roughly 1 million characters in 200 milliseconds, making incremental updates inexpensive when source files change.
Its SAG2 retrieval method builds a local index without embeddings or LLM calls. At query time, it expands the search and applies Jev evidence scoring, which ranks passages according to how strongly they support the query.
External services enter the workflow in specific modes. In-depth semantic retrieval requires a configured Jev service, while generated answers send selected passages to the model provider chosen by the user. Basic indexing remains local.
Read the benchmarks carefully #
The team’s SAG2 paper evaluates evidence recall on 1,000 questions from each benchmark. Recall@10 measures whether the passages needed to answer a question appear among the first 10 results.
This story is for Pro members #
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.