ai-sdk-core-v6_skill

Overview

This skill shows how to build and troubleshoot backend AI using the Vercel AI SDK v5/v6. It focuses on server-side text generation, structured outputs with Zod, and agent workflows with tools and multi-provider support. Practical patterns, error handling, and runtime configuration guidance are included for reliable production use.

How this skill works

The skill explains core APIs: generateText for plain text, generateObject for Zod-validated structured outputs, and streamText for streaming responses. It covers ToolLoopAgent (v6) for agent-driven tool calling and tool approval flows, provider abstraction (openai, anthropic, google), and error types to inspect for robust error handling. Recommendations include runtime callOptions, schema validation, and retry/backoff patterns.

When to use it

• Implement server-side AI features with Vercel AI SDK v5 or v6
• Produce structured JSON outputs validated by Zod
• Build agent workflows that call external tools or require human approval
• Stream model responses for low-latency UX or server-sent events
• Troubleshoot common API and schema validation errors in production

Best practices

• Use generateText for simple prompts and generateObject with Zod for structured responses
• Prefer ToolLoopAgent (v6) for agent orchestration rather than custom loops
• Validate API keys at startup and handle response.error for typed errors
• Implement exponential backoff and retries for rate limits and transient failures
• Use callOptions to adjust model, temperature, and provider per request
• Keep Cloudflare Worker entry modules free of top-level imports to avoid startup overhead

Example use cases

• Backend endpoint that returns typed JSON user summaries using generateObject and a Zod schema
• Agent that queries a database, calls a search tool, and composes an answer with ToolLoopAgent and tool approval
• Streaming chat assistant that uses streamText with onError to gracefully recover and emit partial data
• Multi-provider fallback where requests try openai then anthropic or google based on latency/cost
• Error handling middleware that maps AI_APICallError and AI_NoObjectGeneratedError to actionable responses