Sixian-Li / knowledge-base-kit

# Knowledge Base Kit
**Turn documents into a local, source-backed Markdown knowledge base with Claude
Code or Codex.** Keep the original attachment, a detailed `full.md`, a practical
`summary.md`, and a catalog that agents can read before opening larger files.
简体中文 · Install · Examples · Limitations
## What you get
- **Shared agent skills.** Project-local `process_docs` and `kb` skills for both agents.
- **Document extraction.** Helpers for PDFs, DOCX, static HTML, notebooks, text and images.
- **Isolated image workers.** One image per session, explicit backend selection, bounded
 retries and provenance-aware caches.
- **Checks before filing.** Structure, links, archive folding and TeX validation.
- **A separate workspace.** Your documents stay outside this toolkit repository.
This is an **agent-assisted workflow**, not a hosted service or an automatic
one-command ingestion engine. The main agent writes and reviews the full text and
summary; scripts handle extraction, image workers and checks. Validators cannot
prove factual accuracy or perfect coverage.
## Quick start
You need Python 3.10+, Node.js 20+, Pandoc 3+, and an authenticated Claude Code or
Codex CLI for model-assisted processing. Start from a downloaded or cloned copy
of this repository. Tested versions and platform limits are in
compatibility.
 [code] 
Use `--backend claude` or `--backend both` when initializing for those agents.
Add `--language zh-CN` for Chinese summaries. The installer copies skills and uses
relative project-local links; it does not edit global agent settings or another
knowledge base. It refuses an unrelated nonempty destination.
### Process your first document
Open `my-kb` in your chosen agent and ask:
> Use the local process_docs skill with the Codex backend to process
> inbox/quickstart.md. Prepare the full document and summary, show me the proposed
> category and changed files, and wait for my review before filing.
### Try it offline
For a first look **without an account or model calls**, run this from the toolkit
folder after installing dependencies and Node/Pandoc:
 [code] 
The offline demo extracts the included Markdown, copies a reviewed reference
output and validates it in a fresh workspace. It does not simulate AI generation.
See the reference full document
and summary.
## How it works
Your workspace contains `inbox/`, `catalog.md`, knowledge categories and a private
`.kbkit/` installation. A document has `full.md`, `summary.md` and its original
attachment. A category has `README.md`. Architecture
explains the boundaries; workflows gives the exact steps.
## Configuration and privacy
Edit the workspace's `kb.config.yaml`. Models default to the selected CLI's
default; use a vision-capable model supported by your account. Authentication is
handled by the CLI. This repository contains no provider credentials.
Image workers send images and a short source context to the selected provider.
Text drafting in your main agent also follows that provider's data handl