AI Chat Application

AI Chat Application (Meta-Skill)

Complete guide to building production AI chat applications with the Vercel AI SDK.

Installation

OpenClaw / Moltbot / Clawbot

npx clawhub@latest install ai-chat-application

When to Use

• Building a new AI chatbot or assistant
• Adding AI chat features to an existing app
• Need end-to-end guidance on AI chat architecture

Architecture Overview

┌─────────────────────────────────────────────────────────────┐
│                    Client (React)                            │
├─────────────────────────────────────────────────────────────┤
│  useChat Hook                    │  Data Stream Handler     │
│  - messages, input, submit       │  - Custom data events    │
│  - isLoading, error              │  - Message annotations   │
│  See: vercel-ai-chat-integration │  See: vercel-ai-data-streaming
├─────────────────────────────────────────────────────────────┤
│                    API Route (Next.js)                       │
│  - createDataStreamResponse, streamText, message persistence│
│  See: ai-streaming-routes                                    │
├─────────────────────────────────────────────────────────────┤
│                    Tool System                               │
│  - Toolkit composition, singleton factories, error handling │
│  See: ai-tool-composition, vercel-ai-tool-architecture      │
├─────────────────────────────────────────────────────────────┤
│                    Message Handling                          │
│  - DB ↔ UI conversion, sanitization, annotations           │
│  See: ai-message-handling                                    │
└─────────────────────────────────────────────────────────────┘

Implementation Steps

Step 1: API Route

Create the streaming chat endpoint.

Read: ai/skills/ai-chat/ai-streaming-routes

// app/api/chat/route.ts
export async function POST(request: Request) {
  const { id, messages, modelId } = await request.json();
  
  // Auth, validation, save user message...
  
  return createDataStreamResponse({
    execute: (dataStream) => {
      const result = streamText({
        model: customModel(model),
        system: systemPrompt,
        messages: convertToCoreMessages(messages),
        tools: makeToolkitByName(model.toolkit, params),
        onFinish: async ({ response }) => {
          // Save assistant messages
        },
      });
      result.mergeIntoDataStream(dataStream);
    },
  });
}

Step 2: Chat Component

Build the client-side chat interface.

Read: ai/skills/ai-chat/vercel-ai-chat-integration

// components/chat.tsx
export function Chat({ id, initialMessages }: ChatProps) {
  const { messages, input, handleSubmit, isLoading } = useChat({
    id,
    initialMessages,
    experimental_throttle: 100,
    onError: (error) => toast.error(parseError(error)),
  });

  return (
    <div>
      <Messages messages={messages} isLoading={isLoading} />
      <ChatInput value={input} />
    </div>
  );
}

Step 3: UX Polish

Add auto-scroll and loading states.

const [containerRef, endRef] = useScrollToBottom();

{isLoading && lastMessage?.role === 'user' && <ThinkingMessage />}

Step 4: Tool System

Add function calling capabilities.

Read: ai/skills/ai-chat/ai-tool-composition

// lib/tools/allTools.ts
export function allToolsToolkit(params: ToolkitParams) {
  return {
    ...makeWeb3Tools(params),
    ...makeSearchTools(params),
    getTime: tool({ /* inline tool */ }),
  };
}

Step 5: Message Persistence

Handle message format conversion and persistence.

Read: ai/skills/ai-chat/ai-message-handling

// Convert DB → UI on page load
const uiMessages = convertToUIMessages(dbMessages);

// Sanitize before save
const cleanMessages = sanitizeResponseMessages(response.messages);

Component Skills Reference

NEVER Do

• Never skip error handling in useChat — AI APIs fail; always handle onError
• Never forget to sanitize messages before persistence — Incomplete tool calls corrupt DB
• Never skip throttling for streaming — Causes janky UI with fast token streams
• Never expose raw API errors to users — Parse and show user-friendly messages
• Never block on tool calls without timeout — Tools can hang; set maxSteps and timeouts
• Never save user messages in onFinish — Save immediately before streaming starts
• Never forget maxDuration on API route — Default timeout is too short for AI
• Never skip message annotations — Client needs server-generated IDs for persistence
• Never mutate messages during streaming — Wait for onFinish before state updates

Quick Start Checklist

• Create API route with createDataStreamResponse
• Set up useChat hook in client component
• Add message persistence (save/load)
• Implement auto-scroll behavior
• Add thinking indicator
• Create tool system (if needed)
• Define AI persona via system prompt
• Handle errors gracefully
• Add throttling for smooth streaming