How can I switch between different LLM models with the AI SDK?

The AI SDK provides a unified interface, allowing you to easily swap between different LLM providers and models by simply changing a line of code, without altering your application's core framework.

What is function calling (tool calling) in the context of the AI SDK?

Function calling, or tool calling, allows the LLM to interact with the outside world by invoking predefined functions or tools. You provide the model with a list of tools and it decides when to call them based on the prompt and tool descriptions.

How do I get structured JSON output from the AI SDK?

You can use the `generateObject` function, which allows you to define a specific schema (often with Zod) for the LLM's output, ensuring you receive type-safe, structured data instead of plain text.

How do I build a deep research agent using the AI SDK?

Building a deep research agent involves defining a recursive process where the AI generates search queries, searches the web (e.g., using Exa AI), analyzes the results for relevance, extracts learnings, and formulates follow-up questions, all managed by the AI SDK's primitives.

Can the AI SDK models access real-time information?

While some models have knowledge cutoffs, the AI SDK allows integration with providers like Perplexity, whose models can access real-time information, enabling your applications to provide up-to-date answers.

How does the AI SDK handle inferred parameters for tools?

The AI SDK can leverage the LLM's inference capabilities. For example, if a tool requires latitude and longitude, the model might infer these values from the prompt and tool descriptions without explicit input.

Key Moments

AI Dev 25 x NYC | Aayush Kapoor: Vercel AI SDK: From Fundamentals to Deep Research

DeepLearning.AI

Education4 min read32 min video

Dec 2, 2025|1,685 views|30|1

Save to Pod

Key Moments

TL;DR

Vercel AI SDK simplifies building AI apps with primitives for text gen, tool/function calls, & structured output.

Key Insights

The Vercel AI SDK provides a unified interface in TypeScript for building AI web applications, supporting various frameworks and LLM providers.

Key primitives include `generateText` for basic text generation, a unified `toolCalling` system for integrating external functions or data, and `generateObject` for structured JSON output.

Model and provider swapping is seamless by changing a single line of code, allowing flexibility with over 30 supported integrations.

Function calling enables LLMs to interact with the outside world by defining tools with descriptions, input schemas, and execute functions, allowing for multi-step executions.

`generateObject` with Zod schema validation generates type-safe and parsable structured data, ideal for complex data extraction.

A 'Deep Research' agent example demonstrates chaining these primitives to recursively generate search queries, search the web, analyze results, and compile a markdown report.

INTRODUCTION TO THE VERCEL AI SDK

Aayush Kapoor introduces the Vercel AI SDK as a free, open-source toolkit designed for TypeScript developers to rapidly build AI-powered web applications. It supports popular frameworks like Next.js, Svelte, Vue, and React, along with runtimes like Node.js. The SDK aims to simplify LLM integration, offering a unified interface that allows developers to easily switch between different AI providers and models with minimal code changes. Installation is straightforward via npm or other package managers, setting the stage for practical AI development.

CORE FUNCTIONAL PRIMITIVES

The AI SDK offers fundamental building blocks for AI applications. The `generateText` function is the most basic, enabling direct text generation from an LLM with a specified model (e.g., GPT-4 Mini) and a prompt. A key feature is the unified interface that allows seamless model swapping; developers can switch from OpenAI to Perplexity for real-time data by altering just one line of code, demonstrating flexibility across over 30 supported providers.

TOOL AND FUNCTION CALLING MECHANISMS

Beyond text generation, the SDK supports tool and function calling to extend LLM capabilities. This allows models to interact with the external world by calling predefined functions. Developers define tools with a name, a descriptive purpose for the LLM, an input schema for arguments, and an execute function. The SDK handles parsing the LLM's tool calls and executing the corresponding functions. This enables multi-step processes, like performing calculations or fetching data, and allows the LLM to receive tool execution results back into its context.

ADVANCED TOOL CALLING AND INFERENCE

The function calling system can manage multi-step executions by specifying a `stopWhen` parameter, allowing the LLM to run for more than one step. For instance, it can first call a tool to get information (like weather) and then use that information in a subsequent LLM call. The SDK leverages model inference capabilities, enabling LLMs to infer missing input parameters for tools based on the prompt and tool definitions, such as inferring latitude and longitude for a location-based weather query without explicit user input.

GENERATING STRUCTURED OUTPUTS

For applications requiring machine-readable data, the `generateObject` function is crucial. It allows developers to define a desired JSON schema using libraries like Zod. The LLM then generates data conforming to this schema, ensuring type safety and easier parsing compared to extracting information from plain text. This is demonstrated by requesting a list of hot chocolate spots with their neighborhoods, outputting a structured object that can be directly used in applications.

BUILDING A DEEP RESEARCH AGENT

The session details the construction of a 'Deep Research' agent as a practical application of these primitives. The agent recursively breaks down an initial query into sub-queries, searches the web using an external API like Exa AI, evaluates the relevance of search results, synthesizes learnings, and generates follow-up questions. This recursive process, controlled by a depth parameter, allows for in-depth exploration of a topic, ultimately compiling findings into a markdown report.

AGENT ARCHITECTURE AND RECURSION

The deep research agent's architecture involves functions for generating search queries, searching the web, analyzing results, and generating learnings. The `searchAndProcess` function orchestrates web searching and evaluation. The core `deepResearch` function is designed to be recursive, handling depth and breadth parameters. It iteratively generates queries, processes search results, extracts learnings, and if the depth allows, uses follow-up questions to create new queries, creating a loop for comprehensive research.

FINAL REPORT GENERATION AND NEXT STEPS

Once the recursive research process is complete, a `generateReport` function, utilizing `generateText`, compiles all the accumulated research into a final markdown report. The presentation emphasizes experimentation, encouraging developers to explore the SDK's documentation at a-sdk.dev for more primitives like image generation and streaming. Developers are invited to connect with the presenter on LinkedIn and Twitter for further assistance and to showcase their AI SDK-built applications.

Mentioned in This Episode

●Software & Apps

●Companies

AI SDK Building Blocks & Deep Research Agent

Practical takeaways from this episode

Do This

Use `generateText` for simple text generation.

Leverage `toolCalling` to enable LLMs to interact with external functions.

Configure `stopWhenStepCount` to control multi-step tool execution.

Utilize `generateObject` with Zod schemas for structured JSON output.

Build research agents using a recursive approach with defined depth and breadth.

Explore available primitives like `generateImage` and streaming results.

Refer to the documentation at a-sdk.dev for more examples and features.

Avoid This

Do not expect LLMs to have real-time knowledge without specific tools or models (like Perplexity).

Avoid overly complex prompts without providing clear tool definitions or schemas.

Ensure recursive functions have a base case to prevent infinite loops.

Do not solely rely on LLM output without evaluating relevance, especially for research agents.

Common Questions

The Vercel AI SDK is a free, open-source TypeScript library designed to help developers rapidly build web applications that leverage AI. It offers a unified interface for interacting with various LLM providers and models.