AI

@delightstack/ai adds AI to SvelteKit apps on Cloudflare Workers — chat completions, resumable streaming, vector embeddings, and tool calling, routed through Cloudflare AI Gateway and integrated with the database and realtime layers.

Features

Chat completions — streaming and non-streaming via AI Gateway (Workers AI @cf/ models, OpenAI, Anthropic, or any provider behind the gateway).
Resumable streaming — chunks buffer in the DO and broadcast over WebSocket; clients resume from their last offset after reconnect, with no lost tokens.
Vector embeddings — alarm-based async pipeline with change detection; embeddings generate automatically on create/update.
Tool calling — OpenAI-compatible function calling with streaming delta merging.
Reactive AiClient — Svelte 5 runes for streaming content, usage, errors, and tool calls.

Install

pnpm add @delightstack/ai

Import	Use
`@delightstack/ai/server`	`aiProcessing`, `createAiHandle`, `createAiMessageHandler`
`@delightstack/ai/client`	`AiClient` — reactive Svelte 5 client
`@delightstack/ai/schema`	`defineAiTable`, `defineAiConversationTable`

It builds on database and websocket, and needs an AI binding (Cloudflare AI Gateway) in wrangler.toml.

1. Schema with embeddings

import { defineAiTable } from '@delightstack/ai/schema';

// extends a database table with embedding + embedding_status fields
const articleTable = defineAiTable('article', (schema) => ({
  title: schema.string().searchable(),
  body: schema.string(),
}), { dimensions: 1024 });

2. Server handle

import { createAiHandle } from '@delightstack/ai/server';

export const aiHandle = createAiHandle({
  getAi: (event) => event.platform.env.AI,
  // exposes /api/ai/complete, /api/ai/stream, /api/ai/embed, /api/ai/cancel
});

createAiMessageHandler() wires resume/cancel messages into your WebSocket server so streams survive page transitions and network drops.

3. The reactive client

<script>
  import { AiClient } from '@delightstack/ai/client';
  const ai = new AiClient();

  async function ask() {
    await ai.chat({ messages: [{ role: 'user', content: 'Hello!' }] });
  }
</script>

<button onclick={ask}>Ask</button>
<p>{ai.content}</p>

ai.chat() streams over WebSocket (resumable); ai.complete() returns a full result over HTTP.