ai-sdk-core

Multi-provider LLM workflows use AI SDK v5/v6 to abstract provider differences behind a consistent generateText and streamText API. Define your model once with openai(), anthropic(), or google() and switch providers without rewriting orchestration logic, enabling failover and cost optimization across OpenAI, Anthropic, Google, and Cloudflare Workers AI.

Tool calling with AI SDK involves defining tool schemas, passing them to generateText or streamText, and handling tool use events. Human-in-the-loop approvals intercept tool execution before action, route decisions to a reviewer, and resume the agent workflow based on approval status, enabling safe autonomous agent patterns in production.

AI SDK v5 introduces breaking changes in function signatures and streaming patterns. Migration involves updating generateText and streamObject calls, replacing deprecated provider imports, and refactoring streaming handlers. This Skill provides automated migration guidance and best practices to apply updates systematically across your codebase.

Structured output with Zod uses generateObject and streamObject to enforce schema validation at the API level. AI SDK v5/v6 adapts Zod schemas to provider-specific formats—OpenAI uses JSON schema, Anthropic uses structured output blocks—so you write one schema and it works consistently across all providers.

Streaming with AI SDK uses streamText and streamObject to emit tokens and partial objects as they arrive. Set up event listeners for text chunks or object updates, pipe output to the client, and handle backpressure. This enables real-time AI responses in Next.js, Cloudflare Workers, and Vercel without waiting for full completion.

Document reranking for RAG filters and ranks retrieved chunks by relevance before passing them to the LLM. AI SDK v5/v6 integrates with RAG workflows by accepting reranked context in prompts and tool definitions, reducing token consumption and improving output quality when augmenting generation with external knowledge.