Deployment
Deploy the detection website, these docs, and the same-origin proxy to Cloudflare Pages, or publish them as a static site on GitHub Pages.
bun run build builds both the detection website and these docs. The website goes to web/dist/, and the docs are prerendered and then copied to web/dist/docs/. One static deployment serves the detection website at / and the docs at /docs. The detection website uses HashRouter, so it needs no page route rewrites.
With the connection mode set to this site's proxy (or Auto when a direct call is not allowed), requests go through the same-origin /api/proxy; with Own Worker, your Worker forwards them. All three same-origin entry points import proxyRequest from the single-file Worker worker/main.js:
| File | Runtime |
|---|---|
worker/main.js | The proxy implementation. It can also be deployed on its own as a Cloudflare Worker |
functions/api/proxy.ts | Cloudflare Pages Function |
api/proxy.ts | Vercel Function |
web/scripts/vite-proxy.ts | Vite dev and preview servers |
Cloudflare Pages
Create the project
The Pages project uses the repository root as its root directory and deploys through the Git integration:
| Setting | Value |
|---|---|
| Production branch | main |
| Build command | bun run build |
| Output directory | web/dist |
| Environment variable | BUN_VERSION=1.4.2 |
functions/api/proxy.ts provides the same-origin proxy. wrangler.jsonc pins the compatibility date and enables enable_request_signal, so a client cancellation reaches the upstream.
Check routing and caching
web/public/_routes.json sends only /api/proxy to the Pages Function. The home page, hash routes, docs pages, JS/CSS, fonts, the data manifest, and the zstd chunks are all served directly by Pages static hosting. Responses from /api/proxy are marked no-store.
Do not turn on Cache Everything
Do not add a Cache Everything rule for the whole custom domain. Otherwise it may cache API responses or hold back deployment updates. Pages comes with its own CDN and Tiered Cache. Hash-named JS/CSS and data chunks (/assets/*, /data/chunks/*, /docs/assets/*) are cached in the browser for one year and marked immutable by web/public/_headers. index.html and the data manifest keep the Pages revalidation policy. The long-lived cache depends on web/public/404.html: with this file, Pages answers a missing file with 404. Without it, Pages answers with index.html and status 200, and during a deployment switch-over that HTML is cached as a script for one year, which leaves the page blank. (On 2026-10-01 a missing 404.html left the site blank until the Cloudflare cache was purged and the asset hashes changed.) Do not delete 404.html.
Bind the custom domain
First bind lm.ikale.io in the Pages project, then confirm that Cloudflare DNS points to the *.pages.dev host name of the project. Stop the automatic deployment of the old site only after the static assets on the Pages URL, the JSON/SSE relay of /api/proxy, and the custom domain all pass.
The /api/proxy path of the custom domain lm.ikale.io is served by the proxy Worker route below. The Pages Function serves only *.pages.dev and preview deployments.
Proxy Worker route
wrangler.proxy.jsonc deploys api/proxy.ts as the Worker lm-fingerpoint-proxy and binds the route lm.ikale.io/api/proxy. Worker routes run before Pages, so this path never reaches the Pages Function.
The custom domain does not use the Pages Function because it treats Pages as an origin. Cloudflare replaces a 502 from the origin with its own error page, so the client never sees the upstream's error message. The custom domain also waits at most 100 seconds for the origin's response headers and then returns 524. A Worker route builds the response itself and skips this origin layer.
After you change worker/main.js, push Pages and also run bun run deploy:proxy from the repository root. The API token needs account-level Workers Scripts edit permission and Workers Routes edit permission on the ikale.io zone.
Publish with Wrangler
bun run build:pages
bun run preview:pages # http://127.0.0.1:8788
bun run deploy:pagesIf you created a Direct Upload project, you can also run bun run build:pages:upload and then upload the generated web/dist/ to Pages. This command puts the compiled function into _worker.js. If you upload only the pure static directory that bun run build generates, the same-origin proxy is lost. A Direct Upload project needs the same compatibility date and enable_request_signal as wrangler.jsonc, set in the dashboard.
Static data
At build time, the reference bank, the derived bank, and the binary detector are each compressed with zstd level 19, then split into 4 MiB chunks in web/dist/data/chunks/. Each entry in data/manifest.json has parts and bytes; bytes is the compressed size and is used for download progress. The browser joins and decompresses the files according to the manifest. Detection loads only the data it needs. After the reference bank loads, the website downloads the detector in the background and shows the progress in the top bar. If scoring starts before the download finishes, the computing row and the detect button show that detection is waiting for the detector download. The build fails when the bank and the detector do not match. The largest static file is under the 25 MiB limit of Pages. The CLI package is still built from the redacted derived bank generated at build time and shared_detector.bin. It does not depend on the compressed site data.
GitHub Pages
.github/workflows/pages.yml builds and stores the static artifact on push, and publishes it to GitHub Pages when triggered manually. Before a manual publish, enable GitHub Actions as the Pages source in the repository settings. A project site lives under /<repository>/, so the workflow sets DOCS_BASE to /<repository>/docs, and the docs links, the search index, and llms.txt all use this prefix.
GitHub Pages has no server functions, so this site's proxy /api/proxy does not exist there: with the connection mode Auto or This site's proxy, requests that need the proxy fail. For APIs that allow cross-origin browser calls, choose Direct under Connection in the API configuration to use API mode, including the tokenizer probe. For other APIs, choose Own Worker. The ALLOWED_ORIGINS of the Worker must include the origin of the Pages site.
Vercel
vercel.json builds with bun run build and outputs web/dist. api/proxy.ts serves as the same-origin proxy, runs for at most 300 seconds, and supports cancellation.
Docs site
The docs live in docs/. They are a React Router SPA, and every page is prerendered at build time:
- The HTML and data files of each page.
- The Markdown version of each page,
/docs/llms.mdx/<page>/content.md(/docs/en/llms.mdx/<page>/content.mdin English), for "Copy Markdown" and LLM tools. - The Open Graph image of each page,
/docs/og/<page>/image.png(/docs/en/og/<page>/image.pngin English). - The Open Graph image of the detection website's home page,
/docs/og/detector.png, which the website'sindex.htmluses asog:imageandtwitter:image. - The static search index
api/search, plusllms.txtandllms-full.txt.
DOCS_BASE sets the path prefix of the docs. The default is /docs, and every path above starts with it. SITE_ORIGIN is the public origin of the docs and the detection website, https://lm.ikale.io by default, used for absolute URLs in the Open Graph tags of both and in the Markdown versions; set it to the deployment's own origin when you deploy elsewhere, such as Vercel, a Pages preview, or a fork. For local development, use bun run dev:docs. To build the docs alone, use bun run --cwd docs build.
Last updated on
Tokenizer probe
The probe cancels hidden input with differential counts, makes a Bayesian open-set decision between 62 open tokenizer classes and "unlisted", and picks probe texts by a pairwise upper bound on the error. The result is reference information shown next to the ranking and checked against the model you entered.
Own Worker proxy
Deploy worker/main.js to your own Cloudflare account to get a workers.dev address, so your API key does not pass through this site's proxy.