Custom model

Use any model you can reach over HTTP, or anything you can write in TypeScript, as the model that writes your assistant's answers. You write one function, chat; HelpPuff keeps everything around it: the prompt and its rules, the tools (callbacks, jobs, live chat hand-over, opening hours), citations, guardrails, history and limits.

For an OpenAI-compatible API you do not need code: see Models and providers. Write your own when the API is different, or you want to route, cache, or post-process.

1. A starter file

helppuff scaffold model --use     # writes ./llm.ts and points helppuff.json at it

"model": { "provider": "custom", "module": "./llm.ts", "secrets": ["MY_MODEL_KEY"] }

2. The function

import type { LanguageModel } from '@knowtific/helppuff/sdk';

export default {
  id: 'my-model',
  capabilities: { tools: true },
  async chat(request, ctx) {
    const response = await ctx.fetch('https://api.example.com/v1/chat/completions', {
      method: 'POST',
      headers: { 'Content-Type': 'application/json', Authorization: `Bearer ${String(ctx.env['MY_MODEL_KEY'])}` },
      body: JSON.stringify({ model: request.model, messages: request.messages, max_tokens: request.maxTokens, tools: request.tools }),
      signal: request.signal,
    });
    if (!response.ok) throw new Error(`model ${response.status}`);
    const body = await response.json();
    const message = body.choices[0].message;
    return {
      content: message.content ?? '',
      toolCalls: (message.tool_calls ?? []).map((c) => ({ id: c.id, name: c.function.name, arguments: c.function.arguments })),
      usage: body.usage ? { input: body.usage.prompt_tokens, output: body.usage.completion_tokens } : null,
    };
  },
} satisfies LanguageModel;

What you get (request)

Field	What
model	model.model from helppuff.json (default custom)
messages	OpenAI chat-completions messages: system (HelpPuff's whole prompt, passages included), user, assistant (with tool_calls), tool (results)
tools	The tools offered, OpenAI's function shape. Empty on the last round
maxTokens, temperature, reasoning	The answer's limits. Use what your API supports
onText	Present when the visitor watches the reply arrive: call it with each piece of text to stream (optional)
signal	Aborts when the call takes too long

What you return: content (the text), toolCalls (each { id, name, arguments }, the arguments as JSON text), usage ({ input, output } tokens, or null). HelpPuff runs the tools and calls you again with their results, up to three rounds.

With capabilities: { tools: false }, tool calls your model writes in its text (<tool_call>…, or a JSON object naming a tool) are still read.

What you can use (ctx): ctx.fetch, ctx.env (your secrets and the Worker's bindings), ctx.log(event, data) (never log message text or contact details), ctx.siteId.

3. Secrets

List every variable your file reads under secrets, then store each one:

helppuff secret set MY_MODEL_KEY

helppuff deploy uploads them as encrypted Worker secrets.

4. Deploy and test

helppuff deploy
helppuff model test "Do you do emergency callouts?"

The file is bundled into your Worker by wrangler at deploy (TypeScript included); editing it, or a file next to it, redeploys. It runs in the Workers runtime: fetch, Web APIs and npm packages that work on Workers, no filesystem. The import of @knowtific/helppuff/sdk is for types only, so it deploys without installing anything; npm i -D @knowtific/helppuff gives your editor the types.

A model that throws, or answers an error, gives the visitor "The assistant is busy right now"; helppuff model test shows the real error.