Enhance LLM output by turning your MyST site into a RAG resource with progressive disclosure

This project turns a site built with the MyST Document Engine into a resource that LLMs can index, search, and retrieve. MyST modularizes its content into machine-readable chunks, and this tool returns parts of a MyST site as markdown for the LLM to learn from.

Why this project exists¶

Many LLMs benefit from Retrieval Augmented Generation to improve their outputs. This is especially useful for smaller or open-weights LLMs whose context quickly degrades as it fills up. While you can throw Opus/Fable/Astra at many problems, those models are out of reach for many people. Smaller models need more guidance, but are more accessible and robust to long-term changes in the ecosystem. This project makes them more useful by giving them an efficient way into knowledge bases written with the MyST engine.

The motivation for this project is two-fold:

1. reduce the complexity of reading documentation for LLM agents, to make lower-cost, lower-impact LLMs more effective

2. reduce the necessity of writing documentation for LLM agents instead of writing for people

A design goal is to force no changes on existing documentation. We want to leverage pre-existing community practices to enhance LLM workflows, rather than to make people change their behavior for LLMs.

How it works¶

The MyST AST encodes the content of any document as a modular, machine-readable .json file. It also publishes index files that make it easier for machines to discover the structure and metadata across a site built with MyST.

This project builds a command-line tool that uses this structure so an LLM can pull just the information it needs from a community site, without filling its context with tokens it doesn’t need.

The project has two components:

docslice gives command-line-tool access to sections of any deployed MyST site, for progressive disclosure of documentation. Point it at a page URL and it prints just that section, as markdown. It works by parsing the JSON data every MyST site already publishes.

An agent skill follows the skills-spec and teaches your agent the tool; see Agent skill.

Quick Start¶

Try docslice in a terminal to get the “install-the-myst-markdown-cli” section of the MyST Guide.

npm install -g docslice
docslice get 'https://mystmd.org/guide/quickstart#install-the-myst-markdown-cli'

Learn More¶

• Agent skill - the skill file that teaches agents the tool

• docslice - full guide to installation, commands, and library usage

• Evaluation - performance and design analysis

• Development - setup, conventions, and how to contribute