vercel-ai-sdk（Vercel AI SDK 搜索）

Vercel AI SDK v5 Implementation Guide

When to Use This Skill

Use this skill when:

• Implementing AI chat interfaces with useChat hook
• Creating API routes that generate or stream AI responses
• Adding tool calling / function calling capabilities
• Generating text embeddings for semantic search or RAG
• Migrating from AI SDK v4 to v5
• Integrating Model Context Protocol (MCP) servers
• Working with streaming responses or message persistence

Structured Implementation Workflow

  NEVER accept "Module not found" errors as environment issues
  YOU must install the required packages with the CORRECT package manager

  Common packages needed:
  - ai (core AI SDK)
  - @ai-sdk/openai (OpenAI provider)
  - @ai-sdk/anthropic (Anthropic provider)
  - @modelcontextprotocol/sdk (MCP integration)
  - zod (for tool schemas)
</critical>

  "Code is correct" is NOT enough
  You must achieve FULL PASSING status
  This is what it means to be an autonomous agent
</critical>

⚠️ AUTONOMOUS AGENT MINDSET

You are not just writing code - you are COMPLETING TASKS AUTONOMOUSLY.

This means:

1. ✅ Write correct implementation
2. ✅ Install any required dependencies
3. ✅ Run build and fix ALL errors
4. ✅ Run tests and debug ALL failures
5. ✅ Iterate until EVERYTHING passes
6. ✅ Never make excuses or give up

Common Failure Patterns to AVOID

❌ WRONG: "The code is correct, but the package isn't installed - that's an environment issue" ✅ CORRECT: "Build failed due to missing package - installing it now with npm install [package]"

❌ WRONG: "Tests pass but build fails - not my problem" ✅ CORRECT: "Build is failing - debugging the error and fixing it now"

❌ WRONG: "There's a framework bug, can't fix it" ✅ CORRECT: "Framework error detected - researching the issue, trying workarounds, debugging until I find a solution"

❌ WRONG: "The implementation is complete" (with failing tests) ✅ CORRECT: "Tests are failing - debugging and fixing until they all pass"

Dependency Installation Workflow

When you encounter "Module not found" errors:

1. Detect the package manager FIRST - Check for lockfiles: ls -la | grep -E "lock" # Look for: pnpm-lock.yaml, package-lock.json, yarn.lock, bun.lockb
2. Identify the package from the import statement Error: Cannot find module '@ai-sdk/openai' Import: import {openai} from '@ai-sdk/openai' Package needed: @ai-sdk/openai
3. Install with the CORRECT package manager # If pnpm-lock.yaml exists (MOST COMMON for Next.js evals): pnpm install @ai-sdk/openai # or pnpm add @ai-sdk/openai # If package-lock.json exists: npm install @ai-sdk/openai # If yarn.lock exists: yarn add @ai-sdk/openai # If bun.lockb exists: bun install @ai-sdk/openai
4. Re-run build to verify npm run build # or pnpm run build, yarn build, bun run build
5. Fix any new errors that appear

⚠️ CRITICAL WARNING: Using the WRONG package manager (e.g., npm when the project uses pnpm) will:

• Create a second conflicting lockfile
• Install different versions of dependencies
• Cause dependency version mismatches
• Break the build with cryptic errors like "Cannot read properties of null"

Build Error Debugging Workflow

When build fails:

1. Read the FULL error message - don't skim it
2. Identify the root cause:

- Module not found → Install package - Type error → Fix types - Config error → Check config files - Next.js error → Research, try different approaches

1. Apply the fix
2. Re-run build
3. Repeat until build passes

Test Failure Debugging Workflow

When tests fail:

1. Read the FULL test error - understand what's expected
2. Compare expected vs actual behavior
3. Check your implementation against test assertions
4. Fix the issue in your code
5. Re-run tests
6. Repeat until all tests pass

Success Criteria

Task is ONLY complete when:

• ✅ Build passes (npm run build succeeds)
• ✅ Lint passes (npm run lint succeeds)
• ✅ Tests pass (npm run test succeeds)

NEVER stop at "code is correct" - achieve FULL PASSING status!

⚠️ CRITICAL: Tool Calling API - MUST USE tool() Helper

When implementing tool calling, you MUST use the tool() helper function from the 'ai' package.

❌ WRONG - Plain Object (WILL CAUSE BUILD ERROR)

// DO NOT DO THIS - This pattern is INCORRECT
import { z } from 'zod';

tools: {
  myTool: {
    description: 'My tool',
    parameters: z.object({...}),  // ❌ WRONG - "parameters" doesn't exist in v5
    execute: async ({...}) => {...},
  }
}

This will fail with: Type '{description: string; parameters:...}' is not assignable to type '{inputSchema: FlexibleSchema<any>;...}'

✅ CORRECT - Use tool() Helper (REQUIRED)

// ALWAYS DO THIS - This is the ONLY correct pattern
import { tool } from 'ai';  // ⚠️ MUST import tool
import { z } from 'zod';

tools: {
  myTool: tool({  // ⚠️ MUST wrap with tool()
    description: 'My tool',
    inputSchema: z.object({...}),  // ⚠️ MUST use "inputSchema" (not "parameters")
    execute: async ({...}) => {...},
  }),
}

Tool Calling Checklist

Before implementing any tool, verify:

• Imported tool from 'ai' package: import {tool} from 'ai';
• Wrapped tool definition with tool({...})
• Used inputSchema property (NOT parameters)
• Used zod schema: z.object({...})
• Defined execute function with async callback
• Added description string for the tool

⚠️ CRITICAL: Common v4 to v5 Breaking Changes

1. useChat Hook Changes

❌ WRONG (v4 pattern):

const { messages, input, setInput, append } = useChat();

// Sending message
append({ content: text, role: 'user' });

✅ CORRECT (v5 pattern):

const { messages, sendMessage } = useChat();
const [input, setInput] = useState('');

// Sending message
sendMessage({ text: input });

2. Message Structure

❌ WRONG (v4 simple content):

<div>{message.content}</div>

✅ CORRECT (v5 parts-based):

<div>
  {message.parts.map((part, index) =>
    part.type === 'text' ? <span key={index}>{part.text}</span> : null
  )}
</div>

3. Model Specification

✅ PREFER: String-based (v5 recommended):

import { generateText } from 'ai';

const result = await generateText({
  model: 'openai/gpt-4o',  // String format
  prompt: 'Hello',
});

✅ ALSO WORKS: Function-based (legacy support):

import { openai } from '@ai-sdk/openai';
import { generateText } from 'ai';

const result = await generateText({
  model: openai('gpt-4o'),  // Function format
  prompt: 'Hello',
});

Core API Reference

1. generateText - Non-Streaming Text Generation

Purpose: Generate text for non-interactive use cases (email drafts, summaries, agents with tools).

Signature:

import { generateText } from 'ai';

const result = await generateText({
  model: 'openai/gpt-4o',           // String format: 'provider/model-id'
  prompt: 'Your prompt here',        // User input
  system: 'Optional system message', // Optional system instructions
  tools?: { ... },                   // Optional tool calling
  maxSteps?: 5,                      // For multi-step tool calling
});

Return Value:

{
  text: string;              // Generated text output
  toolCalls: ToolCall[];     // Tool invocations made
  finishReason: string;      // Why generation stopped
  usage: TokenUsage;         // Token consumption
  response: RawResponse;     // Raw provider response
  warnings: Warning[];       // Provider-specific alerts
}

Example:

// app/api/generate/route.ts
import { generateText } from 'ai';

export async function GET() {
  const result = await generateText({
    model: 'anthropic/claude-4-sonnet',
    prompt: 'Why is the sky blue?',
  });

  return Response.json({ text: result.text });
}

2. streamText - Streaming Text Generation

Purpose: Stream responses for interactive chat applications.

Signature:

import { streamText } from 'ai';

const result = streamText({
  model: 'openai/gpt-4o',
  prompt: 'Your prompt here',
  system: 'Optional system message',
  messages?: ModelMessage[],  // For chat history
  tools?: { ... },
  onFinish?: async (result) => { ... },
  onError?: async (error) => { ... },
});

Return Methods:

// For chat applications with useChat hook
result.toUIMessageStreamResponse();

// For simple text streaming
result.toTextStreamResponse();

Example - Chat API Route:

// app/api/chat/route.ts
import { streamText, convertToModelMessages } from 'ai';
import type { UIMessage } from 'ai';

export async function POST(req: Request) {
  const { messages }: { messages: UIMessage[] } = await req.json();

  const result = streamText({
    model: 'openai/gpt-4o',
    system: 'You are a helpful assistant.',
    messages: convertToModelMessages(messages),
  });

  return result.toUIMessageStreamResponse();
}

3. useChat Hook - Client-Side Chat Interface

Purpose: Build interactive chat UIs with streaming support.

Signature:

import { useChat } from 'ai/react';

const {
  messages,        // Array of UIMessage with parts-based structure
  sendMessage,     // Function to send messages (replaces append)
  status,          // 'submitted' | 'streaming' | 'ready' | 'error'
  stop,            // Abort current streaming
  regenerate,      // Reprocess last message
  setMessages,     // Manually modify history
  error,           // Error object if request fails
  reload,          // Retry after error
} = useChat({
  api: '/api/chat',  // API endpoint
  onFinish?: (message) => { ... },
  onError?: (error) => { ... },
});

Complete Example:

'use client';

import { useChat } from 'ai/react';
import { useState } from 'react';

export default function ChatPage() {
  const { messages, sendMessage, status } = useChat();
  const [input, setInput] = useState('');

  const handleSubmit = (e: React.FormEvent) => {
    e.preventDefault();
    if (!input.trim()) return;

    sendMessage({ text: input });
    setInput('');
  };

  return (
    <div>
      <div>
        {messages.map((message) => (
          <div key={message.id}>
            <strong>{message.role}:</strong>
            {message.parts.map((part, index) =>
              part.type === 'text' ? (
                <span key={index}>{part.text}</span>
              ) : null
            )}
          </div>
        ))}
      </div>

      <form onSubmit={handleSubmit}>
        <input
          value={input}
          onChange={(e) => setInput(e.target.value)}
          placeholder="Type a message..."
          disabled={status === 'streaming'}
        />
        <button type="submit" disabled={status === 'streaming'}>
          Send
        </button>
      </form>
    </div>
  );
}

4. Tool Calling / Function Calling

Purpose: Enable AI models to call functions with structured parameters.

Defining Tools:

import { tool } from 'ai';
import { z } from 'zod';

const weatherTool = tool({
  description: 'Get the weather in a location',
  inputSchema: z.object({
    location: z.string().describe('The location to get the weather for'),
    unit: z.enum(['C', 'F']).describe('Temperature unit'),
  }),
  execute: async ({ location, unit }) => {
    // Fetch or mock weather data
    return {
      location,
      temperature: 24,
      unit,
      condition: 'Sunny',
    };
  },
});

Using Tools with generateText/streamText:

// app/api/chat/route.ts
import { streamText, convertToModelMessages, tool } from 'ai';
import { z } from 'zod';
import type { UIMessage } from 'ai';

export async function POST(req: Request) {
  const { messages }: { messages: UIMessage[] } = await req.json();

  const result = streamText({
    model: 'openai/gpt-4o',
    messages: convertToModelMessages(messages),
    tools: {
      getWeather: tool({
        description: 'Get the weather for a location',
        inputSchema: z.object({
          city: z.string().describe('The city to get the weather for'),
          unit: z.enum(['C', 'F']).describe('The unit to display the temperature in'),
        }),
        execute: async ({ city, unit }) => {
          // Mock response
          return `It is currently 24°${unit} and Sunny in ${city}!`;
        },
      }),
    },
  });

  return result.toUIMessageStreamResponse();
}

Multi-Step Tool Calling:

const result = await generateText({
  model: 'openai/gpt-4o',
  tools: {
    weather: weatherTool,
    search: searchTool,
  },
  prompt: 'What is the weather in San Francisco and find hotels there?',
  maxSteps: 5,  // Allow up to 5 tool call steps
});

5. Text Embeddings

Purpose: Convert text into numerical vectors for semantic search, RAG, or similarity.

Signature:

import { embed } from 'ai';
import { openai } from '@ai-sdk/openai';

const result = await embed({
  model: openai.textEmbeddingModel('text-embedding-3-small'),
  value: 'Text to embed',
});

Return Value:

{
  embedding: number[];  // Numerical array representing the text
  usage: { tokens: number };  // Token consumption
  response: RawResponse;  // Raw provider response
}

Example - Embedding API Route:

// app/api/embed/route.ts
import { embed } from 'ai';
import { openai } from '@ai-sdk/openai';

export async function GET() {
  const { embedding, usage } = await embed({
    model: openai.textEmbeddingModel('text-embedding-3-small'),
    value: 'sunny day at the beach',
  });

  return Response.json({ embedding, usage });
}

Batch Embeddings:

import { embedMany } from 'ai';

const { embeddings, usage } = await embedMany({
  model: openai.textEmbeddingModel('text-embedding-3-small'),
  values: [
    'sunny day at the beach',
    'rainy afternoon in the city',
    'snowy mountain landscape',
  ],
});

6. Message Utilities

convertToModelMessages: Converts UI messages from useChat into ModelMessage objects for AI functions.

import { convertToModelMessages } from 'ai';
import type { UIMessage } from 'ai';

export async function POST(req: Request) {
  const { messages }: { messages: UIMessage[] } = await req.json();

  const result = streamText({
    model: 'openai/gpt-4o',
    messages: convertToModelMessages(messages),  // Convert for model
  });

  return result.toUIMessageStreamResponse();
}

7. Model Context Protocol (MCP) Integration

Purpose: Connect to external MCP servers for dynamic tool access.

Example:

// app/api/chat/route.ts
import { experimental_createMCPClient, streamText } from 'ai';
import { StreamableHTTPClientTransport } from '@modelcontextprotocol/sdk/client/streamableHttp.js';

export async function POST(req: Request) {
  const { prompt }: { prompt: string } = await req.json();

  try {
    // Connect to MCP server
    const httpTransport = new StreamableHTTPClientTransport(
      new URL('http://localhost:3000/mcp')
    );

    const httpClient = await experimental_createMCPClient({
      transport: httpTransport,
    });

    // Fetch tools from MCP server
    const tools = await httpClient.tools();

    const response = streamText({
      model: 'openai/gpt-4o',
      tools,
      prompt,
      onFinish: async () => {
        await httpClient.close();  // Clean up
      },
      onError: async () => {
        await httpClient.close();  // Clean up on error
      },
    });

    return response.toTextStreamResponse();
  } catch (error) {
    return new Response('Internal Server Error', { status: 500 });
  }
}

Key Points:

• Use experimental_createMCPClient (note: experimental API)
• Always close the client in onFinish and onError
• Tools are fetched dynamically with httpClient.tools()
• Requires @modelcontextprotocol/sdk package

Model Specification Patterns

String-Based (Recommended for v5)

// Format: 'provider/model-id'
model: 'openai/gpt-4o'
model: 'anthropic/claude-4-sonnet'
model: 'google/gemini-2.0-flash'

Function-Based (Legacy Support)

import { openai } from '@ai-sdk/openai';
import { anthropic } from '@ai-sdk/anthropic';

model: openai('gpt-4o')
model: anthropic('claude-4-sonnet')

Embedding Models

import { openai } from '@ai-sdk/openai';

// Text embeddings use a different method
openai.textEmbeddingModel('text-embedding-3-small')
openai.textEmbeddingModel('text-embedding-3-large')

TypeScript Best Practices

Type Imports

import type {
  UIMessage,           // Message type from useChat
  ModelMessage,        // Message type for model functions
  ToolCall,            // Tool call information
  TokenUsage,          // Token consumption data
} from 'ai';

Strongly Typed Tools

import { tool } from 'ai';
import { z } from 'zod';

// Tool helper infers execute parameter types
const myTool = tool({
  description: 'My tool',
  inputSchema: z.object({
    param1: z.string(),
    param2: z.number(),
  }),
  execute: async ({ param1, param2 }) => {
    // param1 is inferred as string
    // param2 is inferred as number
    return { result: 'success' };
  },
});

API Route Types

// app/api/chat/route.ts
import type { UIMessage } from 'ai';

export async function POST(req: Request): Promise<Response> {
  const { messages }: { messages: UIMessage[] } = await req.json();

  // ... implementation
}

Common Patterns

Pattern 1: Simple Chat Application

Client (app/page.tsx):

'use client';

import { useChat } from 'ai/react';
import { useState } from 'react';

export default function Chat() {
  const { messages, sendMessage, status } = useChat();
  const [input, setInput] = useState('');

  return (
    <div>
      {messages.map((m) => (
        <div key={m.id}>
          <strong>{m.role}:</strong>
          {m.parts.map((part, i) =>
            part.type === 'text' ? <span key={i}>{part.text}</span> : null
          )}
        </div>
      ))}
      <form onSubmit={(e) => {
        e.preventDefault();
        sendMessage({ text: input });
        setInput('');
      }}>
        <input value={input} onChange={(e) => setInput(e.target.value)} />
        <button disabled={status === 'streaming'}>Send</button>
      </form>
    </div>
  );
}

Server (app/api/chat/route.ts):

import { streamText, convertToModelMessages } from 'ai';
import type { UIMessage } from 'ai';

export async function POST(req: Request) {
  const { messages }: { messages: UIMessage[] } = await req.json();

  const result = streamText({
    model: 'openai/gpt-4o',
    system: 'You are a helpful assistant.',
    messages: convertToModelMessages(messages),
  });

  return result.toUIMessageStreamResponse();
}

Pattern 2: Chat with Tools

Server with tool calling:

import { streamText, convertToModelMessages, tool } from 'ai';
import { z } from 'zod';
import type { UIMessage } from 'ai';

export async function POST(req: Request) {
  const { messages }: { messages: UIMessage[] } = await req.json();

  const result = streamText({
    model: 'openai/gpt-4o',
    messages: convertToModelMessages(messages),
    tools: {
      getWeather: tool({
        description: 'Get weather for a city',
        inputSchema: z.object({
          city: z.string(),
        }),
        execute: async ({ city }) => {
          // API call or mock data
          return { city, temp: 72, condition: 'Sunny' };
        },
      }),
      searchWeb: tool({
        description: 'Search the web',
        inputSchema: z.object({
          query: z.string(),
        }),
        execute: async ({ query }) => {
          // Search implementation
          return { results: ['...'] };
        },
      }),
    },
  });

  return result.toUIMessageStreamResponse();
}

Pattern 3: Non-Interactive Generation

// app/api/summarize/route.ts
import { generateText } from 'ai';

export async function POST(req: Request) {
  const { text } = await req.json();

  const result = await generateText({
    model: 'anthropic/claude-4-sonnet',
    system: 'You are a summarization expert.',
    prompt: `Summarize this text:\n\n${text}`,
  });

  return Response.json({ summary: result.text });
}

Pattern 4: Semantic Search with Embeddings

// app/api/search/route.ts
import { embed } from 'ai';
import { openai } from '@ai-sdk/openai';

export async function POST(req: Request) {
  const { query } = await req.json();

  // Generate embedding for search query
  const { embedding } = await embed({
    model: openai.textEmbeddingModel('text-embedding-3-small'),
    value: query,
  });

  // Use embedding for similarity search in vector database
  // const results = await vectorDB.search(embedding);

  return Response.json({ embedding, results: [] });
}

Common Pitfalls and Solutions

Pitfall 1: NOT Using tool() Helper for Tools - ⚠️ CRITICAL

This is the most common and critical mistake. Always use tool() helper!

// ❌ WRONG - Plain object (WILL CAUSE BUILD FAILURE)
import { z } from 'zod';

tools: {
  myTool: {
    description: 'My tool',
    parameters: z.object({      // ❌ Wrong property name
      city: z.string(),
    }),
    execute: async ({ city }) => { ... },
  },
}
// Build error: Type '{ description: string; parameters: ... }' is not assignable

// ✅ CORRECT - Use tool() helper (REQUIRED)
import { tool } from 'ai';      // ⚠️ MUST import tool
import { z } from 'zod';

tools: {
  myTool: tool({                // ⚠️ MUST use tool() wrapper
    description: 'My tool',
    inputSchema: z.object({     // ⚠️ Use inputSchema (not parameters)
      city: z.string(),
    }),
    execute: async ({ city }) => { ... },
  }),
}

Pitfall 2: Using v4 useChat API in v5

// ❌ WRONG - v4 pattern
const { input, setInput, append } = useChat();
append({ content: 'Hello', role: 'user' });

// ✅ CORRECT - v5 pattern
const { sendMessage } = useChat();
const [input, setInput] = useState('');
sendMessage({ text: 'Hello' });

Pitfall 3: Accessing message.content instead of message.parts

// ❌ WRONG - v4 pattern
<div>{message.content}</div>

// ✅ CORRECT - v5 parts-based
<div>
  {message.parts.map((part, i) =>
    part.type === 'text' ? <span key={i}>{part.text}</span> : null
  )}
</div>

Pitfall 4: Not Converting UIMessages for Model

// ❌ WRONG - passing UIMessages directly
const result = streamText({
  model: 'openai/gpt-4o',
  messages: messages,  // UIMessage[] - type error
});

// ✅ CORRECT - convert to ModelMessage[]
const result = streamText({
  model: 'openai/gpt-4o',
  messages: convertToModelMessages(messages),
});

Pitfall 5: Forgetting MCP Client Cleanup

// ❌ WRONG - no cleanup
const httpClient = await experimental_createMCPClient({
  transport: httpTransport,
});
const tools = await httpClient.tools();
const response = streamText({ model, tools, prompt });
return response.toTextStreamResponse();

// ✅ CORRECT - cleanup in callbacks
const response = streamText({
  model,
  tools,
  prompt,
  onFinish: async () => {
    await httpClient.close();
  },
  onError: async () => {
    await httpClient.close();
  },
});

Pitfall 6: Using Wrong Response Method

// ❌ WRONG - using text stream for useChat
return result.toTextStreamResponse();  // Won't work with useChat hook

// ✅ CORRECT - use UI message stream for useChat
return result.toUIMessageStreamResponse();

// ✅ ALSO CORRECT - text stream for non-chat scenarios
// For simple text streaming (not using useChat hook)
return result.toTextStreamResponse();

Pitfall 7: Wrong Embedding Model Method

// ❌ WRONG - using regular model method
const { embedding } = await embed({
  model: openai('text-embedding-3-small'),  // Wrong method
  value: 'text',
});

// ✅ CORRECT - use textEmbeddingModel
const { embedding } = await embed({
  model: openai.textEmbeddingModel('text-embedding-3-small'),
  value: 'text',
});

Migration Checklist (v4 → v5)

When migrating from v4 to v5, update:

• Replace append with sendMessage in useChat
• Remove input, setInput, handleInputChange from useChat destructuring
• Add local state management for input: const [input, setInput] = useState('')
• Update message rendering from message.content to message.parts.map(...)
• Update sendMessage calls to use {text: input} structure
• Verify convertToModelMessages is used in API routes
• Check that toUIMessageStreamResponse() is used (not v4 streaming methods)
• Update tool definitions to use tool() helper with inputSchema
• Consider using string-based model specification ('provider/model-id')
• Update TypeScript types (UIMessage, ModelMessage)

Decision Guide

When implementing AI SDK features, ask:

1. Is this client-side or server-side?

- Client: Use useChat hook - Server: Use generateText or streamText

1. Do I need streaming or non-streaming?

- Streaming chat: streamText + toUIMessageStreamResponse() - Non-streaming: generateText - Simple text stream: streamText + toTextStreamResponse()

1. Do I need tool calling?

- Yes: Define tools with tool() helper and inputSchema (zod) - Pass tools object to generateText or streamText

1. Am I using the correct message format?

- Client (useChat): Returns UIMessage[] with parts property - Server: Convert with convertToModelMessages() to ModelMessage[] - Render messages using message.parts.map(...)

1. Is my model specification correct?

- Prefer string format: 'openai/gpt-4o' - Function format also works: openai('gpt-4o') - Embeddings: openai.textEmbeddingModel('text-embedding-3-small')

1. Do I need embeddings?

- Use embed for single values - Use embedMany for batches - Use textEmbeddingModel() method

Quick Reference

Additional Resources

When in doubt, check the official documentation:

• Main docs: https://ai-sdk.dev/docs
• API reference: https://ai-sdk.dev/docs/reference
• Examples: https://ai-sdk.dev/examples

Remember: AI SDK v5 uses string-based model specification, parts-based messages, sendMessage instead of append, and requires convertToModelMessages in API routes.