AI-friendly output

Docfuse generates machine-readable files together with the site. The build does not call an external model and does not require an API key.

Generated files

- - llms.txt
  - llms-full.txt
  - manifest.json
    index.md
    pages.json
    summaries.json
    code-examples.json
    0001.jsonl
  - index.html
    index.md
  - index.html
    index.md


`llms.txt`	Titles and site URLs selected by `ai.versions`	Crawler entry point
`llms-full.txt`	Full Markdown while within budget; otherwise a pointer to the shard manifest	Compatible entry point for small and large sites
`ai/manifest.json`	Schema version, budgets, totals, partitions, shard paths, sizes, record counts, and SHA-256 fingerprints	Stable discovery for complete corpora
`ai/content/{version}/{locale}/*.jsonl`	Size-bounded content records with `part` / `parts`	Streaming RAG ingestion without a monolithic download
`ai/index.md`	Title-to-Markdown-source list	Generic Markdown index
`ai/pages.json`	Routes, locales, heading trees, tags, owners, frontmatter, and timestamps	Metadata before RAG chunking
`ai/summaries.json`	Deterministic page summaries	Result previews and route filtering
`ai/code-examples.json`	Routes, locales, and fenced code	Code-example search
Per-page `index.md`	Markdown or MDX source used by the build	Page-level retrieval and citation

Summary generation

Docfuse makes no network request for summaries. It uses, in order:

Frontmatter description.
The first body paragraph that is not a heading or code block.
An empty string when no suitable paragraph exists.

Write a precise description for every public page instead of a promotional slogan.

Inclusion rules

Published pages in the current version are included by default.
Historical versions are excluded to avoid multiple outdated answers for the same topic.
Pages with status: draft are not built.
ai: false excludes a page.
tags and owner are preserved in pages.json.

yaml

---title: Internal incident responsedescription: Incident steps for the private networkai: false---

Configuration

export default {  ai: {    llmsTxt: true,    llmsFullTxt: true,    markdownIndex: true,    pageSummaries: true,    codeExamples: true,    chunkSizeBytes: 262144,    llmsFullMaxBytes: 10485760,    llmsFullOverflow: 'manifest',    versions: 'current'  }}

pages.json and the versioned shard manifest are always generated. The other files follow their switches. A content record never exceeds chunkSizeBytes; an unusually large page is split on Unicode code-point boundaries and reconstructed in part order. When the full corpus exceeds llmsFullMaxBytes, manifest keeps llms-full.txt small and points to the lossless shards, while error fails the atomic build. versions: 'all' includes historical snapshots in isolated version/locale partitions; current remains the safer default.

Private documentation

AI files and neighboring Markdown sources are static site output. An access layer for a private site must protect /*.md, /ai/*, /llms*.txt, and search indexes as well as HTML pages.