Vercel has updated its AI SDK, with @ai-sdk/fish-audio@3.0.0 now available, adding a new provider for speech and transcription models. The updates also include changes to the core ai package, @ai-sdk/valibot@3.0.22, @ai-sdk/vercel@3.0.24, @ai-sdk/vue@4.0.54, and @ai-sdk/voyage@2.0.22, among others.
Key Points
- A new Fish Audio provider has been added to the AI SDK, supporting speech and transcription models.
- The ToolLoopAgent now allows
prepareCallcallbacks to read and override the top-level reasoning option. - A
defaultInstructionsMiddlewarehas been introduced for applying default language model instructions while preserving call-level overrides. - Preceding assistant messages are now preserved when regenerating a response.
- Anthropic generations that are spliced are now rejected, though duplicate message start events for the active message are still allowed.
Context
According to Vercel, these updates are part of a series of version packages released through the Changesets release GitHub action. The @ai-sdk/anthropic@4.0.32 package also received specific patch changes related to generation handling. The ai@7.0.54 package includes changes that enhance how language model instructions are managed and how agent calls are prepared.
Why It Matters
These updates offer developers using the Vercel AI SDK new capabilities for integrating speech and transcription, and provide more granular control over agent behavior and language model instructions. The changes aim to improve the flexibility and reliability of AI-powered applications built with the SDK.
What To Do
- Review the @ai-sdk/fish-audio@3.0.0 documentation for integrating speech and transcription models.
- Explore the
ToolLoopAgent'sprepareCallcallbacks to customize reasoning options. - Implement
defaultInstructionsMiddlewareto manage language model instructions effectively. - Note the changes in how Anthropic generations are handled, particularly regarding spliced generations.
