Ask AI for your Docusaurus site

@leuria/docusaurus gives your docs an assistant that answers from your pages, with links to them, on each reader's own AI. You hold no API keys and pay nothing per question.

npm install @leuria/docusaurus

// docusaurus.config.js
export default {
  plugins: [
    ["@leuria/docusaurus", { suggestions: ["How do I get started?", "How do I deploy?"] }],
  ],
}

That's all. The navbar gets an Ask AI button, and the panel it opens knows every page of your docs.

What it adds​

• Ask AI. A panel that answers from your docs. The assistant searches them, reads the pages it needs and links to them. It knows which page the reader has open, and the text they selected. The conversation stays open while the reader moves between pages.
• Search, if you want it (search: true). A ⌘K search by words at once, and by meaning once the reader has an embedding model. It offers to ask the question instead.
• The Connect button, if you want it (connectButton: true). The Ask panel offers it anyway when it's needed.
• Tools for the browser's agents (WebMCP). The same docs tools (search_docs, read_page, list_pages, open_page) are offered to agents built into the browser.

At build time the plugin makes a small index of your docs: each page's title, description and headings (drafts and unlisted pages are left out). Readers download it the first time they ask, never with the page. The assistant reads the pages it needs from your site, like a reader would.

If your site uses Algolia DocSearch, the assistant also searches with it, so it finds any word of your pages. With search: true, the index keeps the pages' text too, so Leuria's search finds any word: a bigger download, for sites with no search of their own.

Options​

For versioned docs, the assistant and the search stay in the version the reader is on; outside a docs page they use the latest.

Next to your search​

The plugin replaces no theme component, so it works next to Algolia DocSearch, a local search plugin, or your own swizzled components. Its buttons are plain html navbar items holding web components, added once by the plugin. The Ask button doesn't take ⌘K; Leuria's search does, and only when you turn it on.

To place the buttons yourself, set askButton: "none" and add them where you like:

navbar: {
  items: [{ type: "html", position: "right", value: "<leuria-ask-button></leuria-ask-button>" }],
}

The elements keep their own styles in a shadow root and follow your site's light and dark theme. The panel opens below your navbar.

What readers see​

Leuria comes first: the panel proposes the reader's own AI (ChatGPT through Codex, Claude Code, a model in LM Studio or Ollama), which searches and reads your docs with tools.

• Without Leuria, Connect explains how to get it, with Download first. Below, in small, the reader may pick another AI for the page instead: their browser's built-in model, where there is one, or your server, if you set server. Nothing is remembered: Leuria is proposed again on the next visit. With fallback: "auto", these answer at once instead.
• Small models can't use tools, so with the browser's model the panel searches your docs itself and gives it the best passages with the question.
• A reader using another AI still sees a one-line offer to use their own.

Questions go to the reader's AI, never to your server (unless you set server and nothing else can answer). The docs index stays in the reader's browser. With Algolia DocSearch, search words also go to Algolia, as with your own search box.

For coding agents: llms.txt​

Leuria's panel serves people. For coding agents, add docusaurus-plugin-llms: it writes llms.txt and a Markdown copy of every page at build time. The two work side by side.

Other sites​

The plugin is a thin layer on @leuria/docs, which works on any site.

At build time, turn your pages into a corpus:

import { buildCorpus } from "@leuria/docs/build"

const corpus = buildCorpus(
  [{ url: "/docs/intro", title: "Introduction", markdown: readFileSync("docs/intro.md", "utf8") }],
  { site: "My docs", light: true }, // light: titles and headings only; pages are read from the site
)
writeFileSync("public/corpus.json", JSON.stringify(corpus))

In the page:

import { leuria, promptAPI, createAI } from "@leuria/client"
import { createDocs, setupDocs } from "@leuria/docs"

const ai = createAI({ providers: [leuria({ app: "My docs" }), promptAPI()] })
const docs = createDocs(ai, {
  corpus: () => fetch("/corpus.json").then((r) => r.json()), // fetched on first use
  navigate: (url) => router.push(url), // optional: your router, instead of a page load
})
setupDocs(docs, { suggestions: ["How do I get started?"] })
docs.expose() // the docs tools for the browser's agents (WebMCP)

<leuria-search-button></leuria-search-button>
<leuria-ask-button></leuria-ask-button>

• openAsk(question?) and openSearch(query?) open the panel and the search from your own buttons.
• docs.search(query) returns hits by words and by meaning, for your own search UI; docs.tools gives the assistant's tools to your own conversation.
• A light corpus reads pages from your site and keeps their <article> (else <main>): the pages must be rendered on the server, not only in the browser. Leave light off to ship the text instead.
• search: algoliaSearch({ appId, apiKey, indexName }) adds your Algolia DocSearch to the search by words (use the public search-only key).
• createDocs options: version (the version the reader is on), meaning: false (words only), indexName and storage (where vectors are kept).
• Set --leuria-ask-top on leuria-ask to open the panel below your header.