Provider: Cloudflare AI Search

"knowledge": { "retrieval": { "type": "ai-search" } }, with any model (Workers AI by default). Older projects' "backend": { "type": "cloudflare" } is converted by helppuff upgrade: AI Search finds the passages, the model writes the answer.

Cloudflare AI Search does the retrieval and generation: it indexes your site and files, and answers with a Workers AI model. Only a Cloudflare account is needed.

Choose it when you already use AI Search, or want Cloudflare to manage the index. Otherwise the default Workers AI backend is usually the better fit: it runs on the free plan, says "not sure" instead of guessing, and has the dashboard's knowledge tools (choose pages, files, facts, a passage tester).

Set up

npx @knowtific/helppuff init --url acme.com --backend cloudflare

init lists the AI Search instances on your account and pre-selects one that already crawls your site (reusing it costs nothing and it is already indexed). Otherwise it creates knowtific-helppuff-<site>.

"backend": { "type": "cloudflare" }                                        // knowtific-helppuff-<site>, created for you
"backend": { "type": "cloudflare", "instance": "acme-website" }            // an instance you already have
"backend": { "type": "cloudflare", "endpoint": "https://search.acme.com" } // a public endpoint, even on another account

Option
instance	—	An instance on your account, by name
endpoint	—	A public AI Search endpoint instead of an instance
model	the instance's	A Workers AI model id, or an AI Gateway alias
maxResults	—	Passages retrieved per question (1–50)

The API token needs AI Search › Edit and AI Search › Run as well as the usual permissions (Configuration).

How your site gets indexed

Your site	What happens
A zone on the same Cloudflare account	AI Search crawls it itself (sitemap mode if the site has one, link-following if not), with a headless browser for pages drawn by JavaScript. It re-syncs every 6 hours.
Anywhere else	HelpPuff crawls up to knowledge.website.maxPages pages (default 50) and uploads them as Markdown. helppuff knowledge sync refreshes them.

Files in knowledge.files (PDF, Word, Markdown, text, HTML, CSV, up to 4 MB each) are uploaded either way. Nothing waits for indexing: deploy goes live at once and answers improve as pages finish (helppuff knowledge status).

On a newly created crawler instance, HelpPuff waits for Cloudflare's first crawl before uploading files, because that first sync drops files uploaded while it runs; then it checks every file is listed.

Notes

• AI Search's chat completions take no tool definitions, so rich messages (option chips, links) use inline markers, which HelpPuff teaches the model and turns into real chips.
• Conversation history (the last 24 turns) is read from HelpPuff's record of the conversation in D1, or kept in KV per session without a database.
• Limits during the beta: 100 instances and 500 crawled pages a day on Workers Free. Workers AI generation is billed as usual.