Working with the Vercel AI SDK

Comprehensive offline documentation for the Vercel AI SDK (ai package), fetched verbatim from ai-sdk.dev and stored in references/. Use these files instead of guessing the API from memory.

How to use this skill

1. Read the bundled references/ first. Files are named by their doc path, with / replaced by __ — e.g. https://ai-sdk.dev/docs/ai-sdk-core/generating-text → references/docs__ai-sdk-core__generating-text.md. Each file carries source/fetched_at/sha256 frontmatter so you know exactly which page and when it was fetched.
2. If something looks stale or missing, fetch the live Markdown — append .md to any ai-sdk.dev page URL (e.g. https://ai-sdk.dev/docs/agents/overview.md) — or query the docs search API: https://ai-sdk.dev/api/search-docs?q=your+query (append .md to returned URLs).
3. grep the references/ directory by topic (generateText, useChat, ToolLoopAgent, a provider name, an error message) to find the right page fast.

Behavioral guidance (when writing AI SDK code)

• Never rely on memory for model IDs. Provider model identifiers change often. Fetch the current list before writing provider code — e.g. the relevant references/providers__*.md page, or the live provider doc.
• Default to the Vercel AI Gateway provider unless the user specifies a direct provider. See references/providers__ai-sdk-providers__ai-gateway.md.
• Check the installed package version. If the project has node_modules/ai/, prefer its docs and match its major version; the AI SDK has meaningful API changes across majors (see the migration guides).
• Prefer streaming + typed output patterns (streamText, streamObject, Zod schemas) where they fit, matching the surrounding codebase.

Reference map

References are grouped by their path prefix:

• docs__getting-started__* — framework quickstarts (Next.js, Node, Svelte, Nuxt, Expo) and library navigation.
• docs__foundations__* — core concepts: providers/models, prompts, tools, streaming, agents.
• docs__ai-sdk-core__* — server-side generation: generateText, streamText, generateObject, streamObject, embeddings, image/speech, tool calling, middleware, provider/model management, error handling, telemetry.
• docs__ai-sdk-ui__* — client framework hooks: useChat, useCompletion, useObject, streaming, message persistence, transport.
• docs__ai-sdk-rsc__* — React Server Components streaming UI (streamUI, generative UI).
• docs__agents__* — building agents, ToolLoopAgent, loop control, memory, workflows, subagents.
• docs__advanced__* — caching, rate limiting, backpressure, stopping streams, rendering UI with language models.
• docs__reference__* — full API reference for AI SDK Core, UI, RSC, and provider utilities.
• docs__troubleshooting__* — common errors and fixes.
• docs__migration-guides__* — version-to-version migration (3.x → 6.0).
• providers__* — first-party and OpenAI-compatible/community provider docs (OpenAI, Anthropic, Google, AI Gateway, Bedrock, Groq, and more).
• cookbook__* — task-oriented recipes for Next.js, Node, and framework-agnostic guides (RAG, multi-modal chat, structured extraction, etc.).

Keeping this current

The references/ directory is regenerated by scripts/update_docs.js, which reads ai-sdk.dev/sitemap.md, filters to the /docs, /providers, and /cookbook sections, and re-fetches each page. A weekly GitHub Action (.github/workflows/update-docs.yml) runs it and commits any changes. To refresh locally: make update-working-with-vercel-ai-sdk.

Note: official live-search skill

Vercel also ships an official, thinner skill (npx skills add vercel/ai, source vercel/ai → skills/use-ai-sdk) that performs live documentation search at runtime instead of bundling docs offline. This skill differs by bundling the full docs locally — prefer it for offline/grep-able access; use the official one if you want only live search.