Image: GitHub
Why it mattersMost knowledge bases send files to a cloud model. SAG-Wiki indexes locally with no model call and traces every answer to the source.
A knowledge base that an agent can search is useful. One that reads a team's notes without sending them to a cloud model is more useful. That second kind is harder to find.
Zleap-AI released SAG-Wiki, a small, local knowledge base that does that. A user points it at a folder on a laptop, and it builds an index. An agent such as Claude Code or Codex can then search across the files and return answers with citations. The repository has 656 stars, 77 forks and an MIT licence. It is ranked 18 on Trendshift's daily list today.
How it works, in plain words
SAG-Wiki's index does not call a language model. It uses only local search. The project's README says it works at about 200 milliseconds per one million characters of text. A full book in that terms is a fraction of a second.
The user downloads a desktop app of about 14 megabytes. The app opens a window that holds a set of named sources, each a folder from the user's machine. The app reads each file as it is added. Markdown, HTML, PDF, Microsoft Office files and EPUB books are covered. Image and audio handling is there too but optional.
From there, two ways to use it.
A person can open the app and ask a question in plain words. The answer holds the passages that match, each with a link to the source file.
An agent gets the same search. It reaches the index through an installable skill for Claude Code, Codex and OpenCode. It can also run the command-line tool on its own, without the desktop app. A Node package covers the npm route.
What is different from an API-first search
SAG-Wiki belongs to a small group of knowledge tools that keep the index on the user's machine. The trade is clear. The index cannot reason over text the way a model can. It returns passages, not summaries. The agent decides what to do with them.
The win is also clear. No file leaves the machine. No model sees the content during indexing. Every answer points at the source line. A user can tell quickly when the agent has quoted something out of context.
Where this helps a working team
A small team with research notes, meeting write-ups, old code spec documents and PDF reports has a classic retrieval problem. The notes are in six places and nobody remembers where. SAG-Wiki is worth trying for that team. Setup takes about an hour on a laptop. The agent gets a search that respects the team's privacy rules.
Zleap-AI is the creator. The project is new. A team adopting it should keep the source folder under revision control. It should also plan to run the full index again when the files move.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
