chore: release

[dgesteves-release]

This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated.

Releases

ondocs@0.12.0

Minor Changes

• #167 a4a7fe2 Thanks @dgesteves! - The Astro, Starlight and VitePress plugins no longer need React. Their dialog now comes prebuilt with its own React, Radix and cmdk, from the new ondocs/embed/standalone, so a site installs ondocs, ai and @ai-sdk/openai and nothing else. Before, a Vue or Astro site without React had to install react, react-dom, @radix-ui/react-dialog and cmdk, or its build failed. ondocs init no longer lists them for these sites, and starlight-ondocs no longer asks for them as peers, so npm no longer installs them with it. The dialog still loads only the first time it is wanted, in a chunk of its own: 92.6 KB gzipped on the Starlight example (93.6 KB before) and 95.6 KB on the VitePress one (96.8 KB before). A site that installed the four packages only for the dialog can remove them. On an Astro site with React islands, the dialog no longer shares their React.

  ondocs/embed/standalone has the same mountAskDialog and openAskDialog as ondocs/embed, for an app with a bundler and no React. On such a site, import openAskDialog from it: ondocs/embed renders the dialog with the site's React, so the build fails without React installed. launcherScript from ondocs/integration now mounts the launcher with it.

Patch Changes

• #163 4b368af Thanks @dgesteves! - Keyword search finds Chinese, Japanese and Korean text. A run of it with no space was one term, which a question matched only where a page had the same run, so every keyword path (MCP search without an embeddingModel, -e none and -e mock indexes, the keyword half of hybrid search) found almost nothing: 怎么安装插件？ was refused on a page headed 安装插件. A run is now split into overlapping pairs of characters, without question words such as 怎么, 如何, どこ and 어떻게, and kana keep their voicing marks (デプロイ was folded to テフロイ). On the docs' five-page test site, a -e none index now answers 7 of 8 Chinese questions with the right page first, up from 1, and 6 of 8 Japanese ones, up from 1, and still refuses all 6 off-topic questions in each language.

  A CJK paragraph longer than a chunk is now cut at 。, ！, ？, ； or ：, and the next chunk repeats the end of the one before, as in other languages; before, it was cut mid-sentence with no overlap. A hard cut never splits a surrogate pair, so a chunk no longer ends or starts with half an emoji.

  The keyword side is built when an index loads, so an index needs no rebuild for it. One with a CJK paragraph longer than a chunk is cut differently now, so ondocs index --check fails on it until you rebuild it, which embeds only the changed chunks again. An -e mock index of CJK pages keeps the vectors of the chunks that did not change, made from the old terms: delete it (and node_modules/.cache/ondocs for a plugin's build) and build it again. Indexes of other languages are unchanged.

• #162 856edfd Thanks @dgesteves! - An endpoint that cannot answer as it is set up now says so, instead of telling every reader to try again:

  • When the model has no API key, or the provider rejects the key (a 401 or a 403, or AI Gateway with no key or OIDC token), the stream's error is now The ask endpoint is misconfigured., followed by a message-metadata part with misconfigured: true. The dialog shows "Answers aren’t set up on this site yet." with no Retry button, where before it showed "The answer could not be generated. Please try again." with a Retry that could not fix it. The endpoint's 500 (an index it could not load, or one embedded with another model) now shows the same words, but keeps Retry: an index that failed to load, such as a remoteIndex the site did not serve, is loaded again on the next request. The new errorMisconfigured label changes the words, and useAsk reports both with reason: 'misconfigured': the first with kind: 'stream', the 500 with kind: 'http'. Other model failures still say to try again.
  • On a page served from your machine, or in a development build, the console now says the endpoint's log has the cause of these errors and of a failed answer, with a link to Troubleshooting.
  • The console's hint for a 404 now says to start the site with ASK_ENDPOINT=http://localhost:8787/api/ask, as the docs do, rather than to point the dialog's endpoint at it, which a deployed site would then post to. It links to the deployment guide instead of the README.
  • ondocs init no longer calls OPENAI_API_KEY optional where one variable serves the build and the functions (Vercel, Netlify, and Next.js on any host). It says that without the key the MCP endpoint works and off-topic questions are refused, but the Ask box cannot answer, and that init --host github-pages answers with Workers AI and no key.
  • A plugin build with embedding: 'mock' now says the index is for local testing only: the Vercel, Netlify and Cloudflare endpoints init writes refuse it with a 500, and the GitHub Pages Worker ignores its vectors.

• #165 30843b9 Thanks @dgesteves! - remoteIndex no longer lets a site that stops answering hold up every question. Before, once the index was due for a check, every request waited for that check, so a hung site held them all until the platform's own fetch gave up.

  • Each fetch of the index is given up after 10 seconds, for the response and the file together, also when a custom fetch ignores the abort signal. The new timeoutSeconds option changes it, up to 2147483, the longest a timer can wait: raise it for a very large index on a slow link. A check that times out keeps the index it has. A first fetch that times out fails its request with an error that names the option, and the next request tries again.
  • The request that finds the index due still waits for the check, now for at most timeoutSeconds, so it gets a redeployed index at once. This also holds on a Worker, which cancels what a request started once it has answered. Requests that arrive during the check no longer wait for it: they are answered from the index in memory.
  • A check whose connection drops part-way through the file now keeps the index it has. Before, it failed every request waiting on it with a 500. A first fetch that drops names the index's URL in its error, as other failed fetches do.

starlight-ondocs@0.12.0

Minor Changes

• #167 a4a7fe2 Thanks @dgesteves! - The Astro, Starlight and VitePress plugins no longer need React. Their dialog now comes prebuilt with its own React, Radix and cmdk, from the new ondocs/embed/standalone, so a site installs ondocs, ai and @ai-sdk/openai and nothing else. Before, a Vue or Astro site without React had to install react, react-dom, @radix-ui/react-dialog and cmdk, or its build failed. ondocs init no longer lists them for these sites, and starlight-ondocs no longer asks for them as peers, so npm no longer installs them with it. The dialog still loads only the first time it is wanted, in a chunk of its own: 92.6 KB gzipped on the Starlight example (93.6 KB before) and 95.6 KB on the VitePress one (96.8 KB before). A site that installed the four packages only for the dialog can remove them. On an Astro site with React islands, the dialog no longer shares their React.

  ondocs/embed/standalone has the same mountAskDialog and openAskDialog as ondocs/embed, for an app with a bundler and no React. On such a site, import openAskDialog from it: ondocs/embed renders the dialog with the site's React, so the build fails without React installed. launcherScript from ondocs/integration now mounts the launcher with it.

Patch Changes

• Updated dependencies [4b368af, 856edfd, 30843b9, a4a7fe2]:
  • ondocs@0.12.0

docusaurus-plugin-ondocs@0.12.0

Patch Changes

• Updated dependencies [4b368af, 856edfd, 30843b9, a4a7fe2]:
  • ondocs@0.12.0