Provider: Google Gemini

"model": { "provider": "gemini" }, with any knowledge (knowledge.retrieval; HelpPuff's own by default). See Models and providers.

This page describes the whole "backend": { "type": "gemini" }, still used for Gemini File Search (Google has no search-only call, so File Search stays with Gemini answering).

Answers with Google Gemini (the Interactions API). Knowledge comes from Gemini File Search (Google chunks, embeds and retrieves), or from HelpPuff's own knowledge base.

Set up

Get a key at aistudio.google.com/apikey, then:

npx @knowtific/helppuff init --url acme.com --backend gemini --api-key "$GEMINI_API_KEY"

"backend": {
  "type": "gemini",
  "model": "gemini-3-flash",
  "apiKey": { "env": "GEMINI_API_KEY" }
}

Option	Default
model	gemini-3-flash	Any Gemini model
apiKey	{ "env": "GEMINI_API_KEY" }	By environment variable name
retrieval	—	"helppuff": HelpPuff's knowledge base instead of File Search
fileSearchStore	—	Filled in by helppuff knowledge sync

Knowledge: two ways

Gemini File Search. helppuff knowledge sync crawls your site (up to knowledge.website.maxPages, default 50), creates a File Search store, and uploads the pages and knowledge.files to it; the store name is written to fileSearchStore. Google keeps the embeddings; the raw files are deleted after 48 hours. Run sync again to refresh. Citations come back as links (for web pages) or a note naming the document.

HelpPuff's knowledge base ("retrieval": "helppuff"): the site is learned on your Cloudflare account as with the default backend, and Gemini only writes the answer from the passages it is given.

Notes

• prompt.md is sent as the system instruction, versioned as usual.
• Conversations continue with previous_interaction_id; the session holds only the id.
• Rich messages are function calls, rendered and never executed. Replies stream.