Vercel AI SDK v5

Production patterns for AI chatbots with persistence, generative UI, and streaming.

Quick Reference

Core Architecture

Message Flow

Client (useChat) → API Route (streamText) → DB (Drizzle)
     ↑                    ↓
     └── UIMessageStream ←┘

Key Imports

// Core
import { streamText, convertToModelMessages, UIMessage } from 'ai'
import { createUIMessageStream, createUIMessageStreamResponse } from 'ai'

// Client
import { useChat } from '@ai-sdk/react'
import { DefaultChatTransport } from 'ai'

// Provider
import { anthropic } from '@ai-sdk/anthropic'

Decision Tree

Streaming Response Type

• Simple text streaming → streamText().toUIMessageStreamResponse()
• Need custom data parts → createUIMessageStream() + writer.write()
• Need both → createUIMessageStream() + writer.merge(result.toUIMessageStream())

Tool Execution Location

• Has server-side data/secrets → Server tool with execute
• Needs client confirmation → Client tool (no execute) + addToolOutput
• Auto-runs on client → Client tool + onToolCall handler

Data Attachment

• Message-level info (tokens, model, timestamps) → messageMetadata
• Dynamic content in message → Data parts with writer.write()
• Temporary status (not persisted) → Transient data parts

Naming Convention

This skill uses agents instead of chats for all database tables, routes, and methods:

• Table: agents (not chats)
• Foreign keys: agentId (not chatId)
• Actions: createAgent(), loadAgent(), deleteAgent()

File Organization

Follow feature-based architecture:

features/
└── agents/
    ├── data/
    │   └── get-agent.ts
    ├── actions/
    │   └── create-agent.ts
    ├── types/
    │   └── message.ts
    └── components/
        ├── server/
        │   └── agent-messages.tsx
        └── client/
            └── chat-input.tsx

db/
├── schema.ts      # Table definitions
├── relations.ts   # Drizzle relations
└── actions.ts     # DB operations (upsert, load, delete)

Essential Patterns

1. Send Only Last Message

// Client
transport: new DefaultChatTransport({
  api: '/api/agent',
  prepareSendMessagesRequest: ({ messages, id }) => ({
    body: { message: messages.at(-1), agentId: id }
  })
})

// Server: Load history, append new message
const previous = await loadAgent(agentId)
const messages = [...previous, message]

2. Persist on Finish

return result.toUIMessageStreamResponse({
  originalMessages: messages,
  onFinish: async ({ messages }) => {
    await upsertMessages({ agentId, messages })
  }
})

3. Handle Disconnects

const result = streamText({ ... })
result.consumeStream() // No await - ensures completion even on disconnect
return result.toUIMessageStreamResponse({ ... })

4. Type-Safe Tools

const tools = { myTool: tool({ ... }) } satisfies ToolSet
type MyTools = InferUITools<typeof tools>
type MyUIMessage = UIMessage<MyMetadata, MyDataParts, MyTools>

Common Gotchas

1. Tool part types are tool-${toolName} not generic tool-call
2. Data parts need id for reconciliation – same ID updates existing part
3. Transient parts only in onData – never in message.parts
4. sendStart: false when using custom start – avoid duplicate start events
5. result.consumeStream() – call without await for disconnect handling