Vercel has updated its AI SDK, introducing the @ai-sdk/fish-audio@3.0.0 provider. This major change adds support for speech and transcription models. The update also includes several patch changes to the core ai@7.0.54 package and other components.
Key Points
- The @ai-sdk/fish-audio@3.0.0 provider has been added to the Vercel AI SDK.
- This new provider supports both speech and transcription models.
- ToolLoopAgent
prepareCallcallbacks can now read and override the top-level reasoning option. - A
defaultInstructionsMiddlewarehas been introduced to apply default language model instructions while preserving call-level overrides. - The SDK now preserves preceding assistant messages when regenerating a response.
- @ai-sdk/anthropic@4.0.32 rejects spliced Anthropic generations but allows duplicate message start events for the active message.
Context
According to Vercel's changelog, these updates are part of routine releases managed by the Changesets release GitHub action. The changes encompass various packages within the AI SDK, including core functionalities and provider-specific adjustments. The @ai-sdk/amazon-bedrock@5.0.45 and @ai-sdk/angular@3.0.54 packages also received dependency updates related to these changes.
Why It Matters
These updates expand the Vercel AI SDK's capabilities by integrating new audio processing models and refining how developers manage model instructions and agent behavior. The enhancements offer more granular control over AI model interactions and improve the reliability of message handling within applications.
What To Do
- Developers can explore the new Fish Audio provider for speech and transcription functionalities.
- Review the updated ToolLoopAgent documentation for
prepareCallcallback options. - Implement
defaultInstructionsMiddlewareto manage language model instructions. - Note the improved handling of assistant messages during response regeneration.
- Check the @ai-sdk/anthropic package for changes in generation handling.
