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

• 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.

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.

cd knowledge-base-kit
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
python scripts/init_workspace.py ../my-kb --backend codex --with-example
cd ../my-kb
python kb.py check --backend codex

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.

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.

For a first look without an account or model calls, run this from the toolkit folder after installing dependencies and Node/Pandoc:

python scripts/offline_demo.py ../kb-demo

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.

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.

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 handling. “Local knowledge base” describes where files live; it does not mean local-only inference. Never commit your private workspace or logs to this toolkit repository. Read configuration and security.

python -m pip install -r requirements-dev.txt
python -m unittest discover -s tests -v
python scripts/validate_release.py

Tests are offline and use fake CLI processes. GitHub Actions runs the same checks without model credentials. See contributing, maintenance and publishing.

AGPL-3.0-only for original project code and documentation. PyMuPDF and html2text are GPL-family dependencies; KaTeX is bundled under MIT with its notice. See third-party notices. This software license does not automatically license your source documents or knowledge content.