vercel-ai-sdk-expert – skill Lập trình cho agent AI

Vercel AI SDK Expert

You are a production-grade Vercel AI SDK expert. You help developers build AI-powered applications, chatbots, and generative UI experiences primarily using Next.js and React. You are an expert in both the ai (AI SDK Core) and @ai-sdk/react (AI SDK UI) packages. You understand streaming, language model integration, system prompts, tool calling (function calling), and structured data generation.

When to Use This Skill

• Use when adding AI chat or text generation features to a React or Next.js app
• Use when streaming LLM responses to a frontend UI
• Use when implementing tool calling / function calling with an LLM
• Use when returning structured data (JSON) from an LLM using generateObject
• Use when building AI-powered generative UIs (streaming React components)
• Use when migrating from direct OpenAI/Anthropic API calls to the unified AI SDK
• Use when troubleshooting streaming issues with useChat or streamText

Version and execution scope

The code below uses legacy AI SDK APIs (ai/react, toDataStreamResponse, parameters, maxSteps). It is not a current-version starter: identify the installed ai and provider package versions, use the matching migration/reference docs, and keep server/client stream protocols compatible. Validate incoming messages, authenticate and rate-limit routes, keep provider keys server-only, bound token/tool budgets and authorize tool side effects independently of model requests. Plan timeout limits vary; inspect the actual deployment instead of assuming 10–15 seconds.

Core Concepts

Why Vercel AI SDK?

The Vercel AI SDK is a unified framework that abstracts away provider-specific APIs (OpenAI, Anthropic, Google Gemini, Mistral). It provides two main layers:

1. AI SDK Core (ai): Server-side functions to interact with LLMs (generateText, streamText, generateObject).

2. AI SDK UI (@ai-sdk/react): Frontend hooks to manage chat state and streaming (useChat, useCompletion).

Server-Side Generation (Core API)

Basic Text Generation

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

// Returns the full string once completion is done (no streaming)
const { text, usage } = await generateText({
  model: openai("gpt-4o"),
  system: "You are a helpful assistant evaluating code.",
  prompt: "Review the following python code...",
});

console.log(text);
console.log(`Tokens used: ${usage.totalTokens}`);

Streaming Text

// app/api/chat/route.ts (Next.js App Router API Route)
import { streamText } from 'ai';
import { openai } from '@ai-sdk/openai';

// Allow streaming responses up to 30 seconds
export const maxDuration = 30;

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

  const result = streamText({
    model: openai('gpt-4o'),
    system: 'You are a friendly customer support bot.',
    messages,
  });

  // Automatically converts the stream to a readable web stream
  return result.toDataStreamResponse();
}

Structured Data (JSON) Generation

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

const { object } = await generateObject({
  model: openai('gpt-4o-2024-08-06'), // Use models good at structured output
  system: 'Extract information from the receipt text.',
  prompt: receiptText,
  // Pass a Zod schema to enforce output structure
  schema: z.object({
    storeName: z.string(),
    totalAmount: z.number(),
    items: z.array(z.object({
      name: z.string(),
      price: z.number(),
    })),
    date: z.string().describe("ISO 8601 date format"),
  }),
});

// `object` is automatically fully typed according to the Zod schema!
console.log(object.totalAmount); 

Frontend UI Hooks

useChat (Conversational UI)

// app/page.tsx (Next.js Client Component)
"use client";

import { useChat } from "ai/react";

export default function Chat() {
  const { messages, input, handleInputChange, handleSubmit, isLoading } = useChat({
    api: "/api/chat", // Points to the streamText route created above
    // Optional callbacks
    onFinish: (message) => console.log("Done streaming:", message),
    onError: (error) => console.error(error)
  });

  return (
    <div className="flex flex-col h-screen max-w-md mx-auto p-4">
      <div className="flex-1 overflow-y-auto mb-4">
        {messages.map((m) => (
          <div key={m.id} className={`mb-4 ${m.role === 'user' ? 'text-right' : 'text-left'}`}>
            <span className={`p-2 rounded-lg inline-block ${m.role === 'user' ? 'bg-blue-500 text-white' : 'bg-gray-200'}`}>
              {m.content}
            </span>
          </div>
        ))}
      </div>
      
      <form onSubmit={handleSubmit} className="flex gap-2">
        <input
          value={input}
          onChange={handleInputChange}
          placeholder="Say something..."
          className="flex-1 p-2 border rounded"
          disabled={isLoading}
        />
        <button type="submit" disabled={isLoading} className="bg-black text-white p-2 rounded">
          Send
        </button>
      </form>
    </div>
  );
}

Tool Calling (Function Calling)

Tools allow the LLM to interact with your code, fetching external data or performing actions before responding to the user.

Server-Side Tool Definition

// app/api/chat/route.ts
import { streamText, tool } from 'ai';
import { openai } from '@ai-sdk/openai';
import { z } from 'zod';

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

  const result = streamText({
    model: openai('gpt-4o'),
    messages,
    tools: {
      getWeather: tool({
        description: 'Get the current weather in a given location',
        parameters: z.object({
          location: z.string().describe('The city and state, e.g. San Francisco, CA'),
          unit: z.enum(['celsius', 'fahrenheit']).optional(),
        }),
        // Execute runs when the LLM decides to call this tool
        execute: async ({ location, unit = 'celsius' }) => {
          // Demo only: replace with an authorized real data integration.
          return `No live weather lookup was performed for ${location}.`;
        },
      }),
    },
    // Allows the LLM to call tools automatically in a loop until it has the answer
    maxSteps: 5, 
  });

  return result.toDataStreamResponse();
}

UI for Multi-Step Tool Calls

When using maxSteps, the useChat hook will display intermediate tool calls if you handle them in the UI.

// Inside the `useChat` messages.map loop
{m.role === 'assistant' && m.toolInvocations?.map((toolInvocation) => (
  <div key={toolInvocation.toolCallId} className="text-sm text-gray-500">
    {toolInvocation.state === 'result' ? (
      <p>✅ Fetched weather for {toolInvocation.args.location}</p>
    ) : (
      <p>⏳ Fetching weather for {toolInvocation.args.location}...</p>
    )}
  </div>
))}

Best Practices

• ✅ Do: Use openai('gpt-4o') or anthropic('claude-3-5-sonnet-20240620') format (from specific provider packages like @ai-sdk/openai) instead of the older edge runtime wrappers.
• ✅ Do: Provide a strict Zod schema and a clear system prompt when using generateObject().
• ✅ Do: Set maxDuration = 30 (or higher if on Pro) in Next.js API routes that use streamText, as LLMs take time to stream responses and Vercel's default is 10-15s.
• ✅ Do: Use tool() with comprehensive description tags on Zod parameters, as the LLM relies entirely on those strings to understand when and how to call the tool.
• ✅ Do: Enable maxSteps: 5 (or similar) when providing tools, otherwise the LLM won't be able to reply to the user *after* seeing the tool result!
• ❌ Don't: Forget to return result.toDataStreamResponse() in Next.js App Router API routes when using streamText; standard JSON responses will break chunking.
• ❌ Don't: Blindly trust the output of generateObject without validation, even though Zod forces the shape — always handle failure states using try/catch.

Troubleshooting

Problem: The streaming chat cuts off abruptly after 10-15 seconds.

Solution: The serverless function timed out. Add export const maxDuration = 30; (or whatever your plan limit is) to the Next.js API route file.

Problem: "Tool execution failed" or the LLM didn't return an answer after using a tool.

Solution: streamText stops immediately after a tool call completes unless you provide maxSteps. Set maxSteps: 2 (or higher) to let the LLM see the tool result and construct a final text response.

Limitations

• Use this skill only when the task clearly matches the scope described above.
• Do not treat the output as a substitute for environment-specific validation, testing, or expert review.
• Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are

… (nội dung đầy đủ có trong gói tải về)