AI features

Kalayaan’s AI features run on Workers AI — no separate API key or third-party account to set up. Turn them on in cms.config.ts:

export default defineConfig({
  // ...
  ai: {
    enabled: true,
    features: ["alt-text", "translate", "editorial-assist"],
  },
});

features picks which capabilities are active: "alt-text", "translate", "editorial-assist", and "semantic-search" (see the free-tier note below — this one is different from the rest).

What each feature does

Alt-text — generates alt text for uploaded media automatically.
Translate — translates a document’s content into another configured locale.
Editorial assist — a family of per-field actions: improve writing, summarize, and generate an SEO title/description. Attach one to a specific field with aiEnrich:
```
body: field.richText({
  aiEnrich: { action: "improve" },
}),
excerpt: field.text({
  aiEnrich: { action: "summarize", dependency: "body" },
}),
```
dependency names the field to read source text from (omit it to have the action rewrite the field in place). This renders as a small inline “Generate with AI” control right on that field in the admin editor — not a generic sidebar panel — so it’s always obvious which part of the document an AI action affects.

Semantic search (the one paid feature)

Turning on "semantic-search" adds a public search endpoint:

GET /api/v1/search?q=<query>&collection=<name>&locale=<locale>&limit=20

Only q is required. Semantic search needs Vectorize, which is on Cloudflare’s paid Workers plan — everything else in Kalayaan runs on the permanent free tier. To keep the free path free by default:

kalayaan doctor and kalayaan deploy print a heads-up the moment semantic-search is enabled, so you always know before you deploy.
Without Vectorize, /api/v1/search still works — it falls back to a plain SQL text-match scan, just without semantic ranking.

Choosing models

Each capability runs on a specific Workers AI model, with a sane default for each — you don’t need to pick one to get started. To override any of them, add a models block:

ai: {
  enabled: true,
  features: ["editorial-assist"],
  models: {
    text: "@cf/meta/llama-3.3-70b-instruct-fp8-fast", // improve/summarize/seo — bigger, higher quality
  },
},

Key	Used for	Default
`text`	improve, summarize, SEO title/description	`@cf/meta/llama-3.1-8b-instruct-fast`
`vision`	alt-text generation	`@cf/meta/llama-3.2-11b-vision-instruct`
`translate`	the translate action	`@cf/meta/m2m100-1.2b`
`embed`	semantic search	`@cf/baai/bge-m3`

Any key you omit falls back to the default — so models: { text: "..." } only changes the text model and leaves the rest alone. Pick a bigger model (like llama-3.3-70b-instruct-fp8-fast) for higher-quality writing help, or a smaller one to spend fewer neurons.

Cloudflare periodically deprecates older model versions — check the Workers AI model catalog if a request starts failing. Kalayaan’s own defaults are updated when that happens, so staying on the default (by omitting models entirely) is the lowest-maintenance choice.

embed is special: if you override it, you must also set embedDimensions to that model’s output size, since the Vectorize index is provisioned with that exact dimension:

models: {
  embed: "@cf/baai/bge-large-en-v1.5",
  embedDimensions: 1024,
},

No AI features enabled?

Everything above is opt-in — leave ai unset (or enabled: false) and none of it runs. The rest of Kalayaan (content modeling, roles, media, deployment) needs no AI configuration at all.