vendored-docs

docvault

Vendor library documentation locally for AI-assisted development. Converts Rust crates (via rustdoc JSON) and Python packages (via runtime introspection) into structured Markdown API references. Zero external dependencies.

Why

LLMs hallucinate API signatures. Vendored docs don't. By keeping accurate, machine-generated API references in your repo, your AI assistant reads real signatures and real docstrings instead of guessing.

Each guide has two tiers:

• Top: Curated patterns, gotchas, and project-specific usage (written by you or your AI)
• Bottom: Full API reference, machine-generated from the source of truth

Install

Add the marketplace and install the plugin from within Claude Code:

/plugin marketplace add zeapo/docvault
/plugin install vendored-docs@docvault

This gives Claude the skill to generate, find, and use vendored docs automatically. You can install at user scope (default, works across all your projects) or project scope (--scope project, shared with collaborators).

After installing, run /reload-plugins if you're in an active session.

Prompting tips

The plugin triggers automatically on phrases like "vendor docs for", "how do I use", or "what's the API for". But you can also be direct:

vendor docs for tokio
document the jax package
look up crate serde

Once docs exist in your repo, Claude checks them before reaching for the internet. You don't need to tell it to — the skill handles that. If you want to add curated patterns on top of a generated reference, just ask:

add common patterns to the tokio docs
add gotchas to the jax guide

Quick Start (manual)

If you prefer to run the scripts directly without the plugin, they live in plugin/skills/vendored-docs/scripts/:

Rust crates

python3 path/to/rustdoc-json-to-md.py datafusion
python3 path/to/rustdoc-json-to-md.py iceberg-datafusion 0.9.0

Requires rustup toolchain install nightly (stable stays your default).

Python packages

python3 path/to/pydoc-to-md.py jax
python3 path/to/pydoc-to-md.py optax --depth 1
python3 path/to/pydoc-to-md.py jax --include numpy --include random

Must run with a Python environment where the package is installed.

Output

Guides are written to docs/vendored/{lang}/{name}-{version}.md — Rust crates go to docs/vendored/rust/ and Python packages to docs/vendored/python/. This avoids name collisions between packages in different languages. Re-running the script updates the API reference while preserving any curated content you've added above the # Full API Reference separator.

Example output structure

docs/vendored/
  rust/
    datafusion-52.4.0.md      # 6,000+ lines — full Rust API + curated patterns
    iceberg-datafusion-0.9.0.md
  python/
    jax-0.9.2.md              # 48,000+ lines — full Python API
    optax-0.2.7.md

CLI Reference

rustdoc-json-to-md.py

rustdoc-json-to-md.py CRATE [VERSION]   Full pipeline
rustdoc-json-to-md.py CRATE --skip-build Reuse existing JSON
rustdoc-json-to-md.py --json FILE        Convert existing JSON
rustdoc-json-to-md.py CRATE --only-index Compact index
rustdoc-json-to-md.py CRATE --stdout     Write to stdout
rustdoc-json-to-md.py CRATE -o PATH      Custom output path

pydoc-to-md.py

pydoc-to-md.py PKG [VERSION]        Full pipeline
pydoc-to-md.py PKG --depth N         Sub-module depth (default: 2)
pydoc-to-md.py PKG --include MOD     Only specific sub-modules
pydoc-to-md.py PKG --only-index      Compact index
pydoc-to-md.py PKG --stdout          Write to stdout
pydoc-to-md.py PKG -o PATH           Custom output path

How It Works

Rust

1. Runs cargo +nightly rustdoc to generate rustdoc JSON — a structured representation of every public type, method, trait, and doc comment
2. A Python script converts the JSON to clean Markdown, rendering type signatures back to Rust syntax
3. Output includes full method docs with inline examples (from /// doc comments)

Python

1. Imports the package and walks it via inspect.getmembers()
2. Extracts signatures via inspect.signature() (falls back to docstring parsing for C extensions)
3. Collects complete docstrings including Args/Returns/Examples sections

Both scripts use only Python stdlib — zero external dependencies.

License

MIT