vercel-ai-sdk

Streaming chat with tool calls uses the Vercel AI SDK's useChat hook to manage chat state and stream responses, paired with tool definitions that define callable functions. The hook handles client-server communication, renders messages in real time, and coordinates tool execution and outputs within the chat flow.

Multi-step reasoning in chat is supported through the Vercel AI SDK's message parts structure, which separates text, tool outputs, and reasoning steps. Server-side streaming via streamText and toUIMessageStreamResponse enables each reasoning step to stream independently, creating a transparent, step-by-step response.

File attachments integrate into streaming chat through the Vercel AI SDK's message parts architecture, which treats attachments as structured components alongside text and tool outputs. The useChat hook manages attachment handling, and server streaming preserves attachment context across tool calls and reasoning steps.

Client-side tool execution runs functions in the browser and returns results to the chat, while server-side tool execution runs on the server and streams results back. The Vercel AI SDK supports both patterns; server-side streaming via streamText is type-safe and coordinates complex tool sequences, while client-side is faster for simple operations.

Server streaming is optional but recommended for responsive chat experiences. The Vercel AI SDK supports streaming via streamText and toUIMessageStreamResponse to send partial responses in real time, improving perceived performance. Non-streaming is possible but delivers full responses after computation completes.

Rich messages are structured using the Vercel AI SDK's UIMessage format, which composes parts: text fragments, tool invocations with states, reasoning traces, and file references. The useChat hook renders and manages these parts, and toUIMessageStreamResponse serializes them for streaming, enabling granular control over message composition and display.