LLM-friendly output

LLM-friendly output#

2026-06-10

4 min read time

rocm-docs-core supports three features that make documentation more accessible to AI coding assistants and agents:

  • A per-page download button that lets users and agents copy page content directly into an AI context window.

  • A curated llms.txt index file that AI agents can discover and use as an entry point to the documentation.

  • A generated llms-full.txt file that aggregates prose content from all source files into a single plain-text document.

All three features are opt-in and configured in conf.py.

Per-page download button#

The Sphinx Book Theme includes a built-in download button that lets readers download the current page as Markdown or RST. Enable it by setting use_download_button in html_theme_options:

html_theme_options = {
    "use_download_button": True,
}

Once enabled, a download icon appears on each page. Clicking it downloads the source file for that page, which users and agents can paste directly into an AI context window.

llms.txt#

llms.txt is a curated index of your documentation pages, following the llmstxt.org convention for AI agent documentation discovery. It is also compatible with Read the Docs llms-txt support.

Creating the index file#

Create a llms.txt file in your documentation source directory (the same directory as conf.py). Write it as a Markdown document with a project description followed by links to your key pages:

# Project name

> One-sentence description of the project.

## User guide

- [Page title](https://your-docs-url/page-slug): Short description of the page.
- [Another page](https://your-docs-url/another-page): Short description.

Publishing the index file#

Add llms.txt to html_extra_path in conf.py so Sphinx copies it to the output root:

html_extra_path = ["llms.txt"]

After a successful build, llms.txt is available at {project_url}/llms.txt.

Header behavior in the full-text file#

If llms.txt is present in the source directory, it is also used as the header section of llms-full.txt. If no llms.txt is present, the Sphinx project name from conf.py is used as the header instead.

llms-full.txt#

llms-full.txt is generated by rocm-docs-core after each successful build. It aggregates prose content from all .md and .rst source files into a single plain-text document.

Enabling generation#

Set rocm_docs_generate_llms_full = True in conf.py:

rocm_docs_generate_llms_full = True

The file is not generated if the build fails.

Output location#

llms-full.txt is written to the Sphinx output directory alongside the built HTML. For a standard build this is docs/_build/html/llms-full.txt, making it available at {project_url}/llms-full.txt.

Content filtering#

The generator filters source files to keep only meaningful prose:

  • Files with fewer than 10 prose lines are skipped entirely, excluding navigation-only or boilerplate pages.

  • Files in _build, _static, _templates, .git, and .venv directories are excluded.

  • Markup noise is stripped from included files: directive syntax, anchor labels, table separator rows, raw HTML blocks, and RST section underlines.

  • Code blocks are preserved.

Each included file is appended as a section separated by ---, with its path relative to the source directory used as the section heading.

Example configuration#

The following example enables all three features together:

html_theme_options = {
    "use_download_button": True,
}

html_extra_path = ["llms.txt"]

rocm_docs_generate_llms_full = True