llms-full.txt

llms-full.txt is the same idea as llms.txt, inverted. Instead of an index of links and descriptions, it contains the full markdown body of every documentation page concatenated into a single file. The header and per-page metadata still live at the top of each block, but the body is included verbatim.

Flux serves one at docs.fluxhq.dev/llms-full.txt.

Shape of the file

# Flux Docs — Full Corpus

---
# Defining workflows

_Source: https://docs.fluxhq.dev/build/workflows/defining-workflows_

A workflow is a Python function decorated with `@workflow`...
[full page body here]

---
# Workflow inputs

_Source: https://docs.fluxhq.dev/build/workflows/inputs_

[full page body here]

...

A document-level header, then one ----delimited block per page. Each block carries its title, source URL, and complete body. Page order matches llms.txt — lexicographic by slug.

Size

The file is large. As of this writing the corpus is roughly 210 pages across eleven sections, and the file weighs in at several megabytes. That is too large to paste into most chat windows, and it will not fit in a single context for older or smaller models. Use it where bulk is the point: ingestion jobs, embedding pipelines, evaluation harnesses that need the whole surface area in one fetch.

When to use which

The two files exist for different jobs.

• Use llms.txt when an LLM needs to navigate the docs interactively. The index is small enough to fit inside any modern model’s context, the descriptions are enough for the model to decide what to fetch, and per-page fetches keep the working context lean. This is the right pattern for chat assistants and IDE tools.
• Use llms-full.txt when you want the whole corpus in one place. Chunk and embed for RAG. Run an offline evaluation. Build a custom search index. Train a domain-tuned retriever. The full file is the canonical, deploy-time snapshot of every page — easier to consume than scraping the site or walking the GitHub repo.

The rule of thumb: if a human-in-the-loop will steer the model, llms.txt is what you want. If a batch job will consume everything without supervision, reach for llms-full.txt.

How Flux generates it

Same generator as llms.txt. site/scripts/generate-llms-txt.mjs walks the content collection, reads each MDX file’s frontmatter and body, and emits both files into dist/ at build time. The full-corpus variant includes the post-frontmatter body verbatim — MDX components survive as JSX text, code blocks and admonitions included. Downstream chunking should be aware that some lines are component invocations rather than prose.

If the size becomes a problem for a specific consumer, the right move is usually to filter llms-full.txt down to the sections that matter (Build + Reference, say) rather than scraping the rendered site. The file is plain text and per-page sections are clearly delimited.