Vercel has released updates to its AI SDK, including changes to how generateText responses are handled when tool choices are involved. The update, noted in the @ai-sdk/xai@3.0.130 and @ai-sdk/zai@2.0.2 changelogs, introduces a mechanism to reject responses that do not meet specified tool choice criteria.
Additionally, the @ai-sdk/amazon-bedrock@4.0.166 package received an update to support a model family setting for embeddings, enabling the use of Amazon Resource Names (ARNs). Further updates in @ai-sdk/tui@1.0.86 and @ai-sdk/svelte@5.0.85 address image generation cost summation and Anthropic reasoning budgets.
Key Points
- The Vercel AI SDK now rejects
generateTextresponses that do not satisfy a required or specifically selected tool choice. - Normalized response content is exposed on
ToolChoiceViolationErrorfor opt-in recovery. - Tool approval secrets are now allowed in
ToolLoopAgentsettings andprepareCall. - The
@ai-sdk/amazon-bedrock@4.0.166package adds a model family setting for embeddings to support ARNs. - Assistant messages are omitted when only cache points remain after unsigned reasoning is filtered.
- The SDK now sums Gateway image-generation costs across split requests.
- Anthropic reasoning budgets are enabled for application inference profile ARNs.
Context
According to the Vercel AI SDK changelogs, these updates are part of ongoing maintenance and feature enhancements across various SDK packages. The changes aim to improve the reliability of tool usage, enhance provider-specific functionalities, and refine cost tracking for AI operations.
Why It Matters
These updates provide developers with more robust control over tool interactions in AI applications and improve compatibility with cloud provider-specific features. The ability to recover from tool choice violations and better manage Amazon Bedrock embeddings can streamline development workflows and enhance application stability.
What To Do
- Review the
ToolChoiceViolationErrordocumentation for details on opt-in recovery. - Developers using Amazon Bedrock should explore the new model family setting for embeddings to leverage ARN support.
- Note the changes to image generation cost summation in the Gateway package.
- If using Anthropic models, be aware of the enabled reasoning budgets for application inference profile ARNs.
Keep Exploring
/atlas/claude-family /atlas/amazon-bedrock /techniques/output-schema /techniques/system-user-separation
