Vercel has updated its AI SDK, with ai@7.0.94 introducing several changes. These updates include new features for batch processing and streaming text, alongside improvements in handling image results and provider-specific prompt blocks.
Key Points
- ai@7.0.94 now supports per-request models within batch operations.
- The streamText function in the AI SDK enforces tool choices.
- The SDK includes retry logic for unclassified empty image results.
- Retry-attempt accounting is preserved during the retry process.
- A provider-independent result retryability classification has been added.
- Google and Google Vertex prompt blocks are now marked as terminal.
- Dependencies for @ai-sdk/provider@4.0.11, @ai-sdk/gateway@4.0.76, @ai-sdk/provider-utils@5.0.37, and @ai-sdk/alibaba@2.0.42 have been updated.
Context
According to Vercel, these changes were released via the Changesets release GitHub action. The updates affect various components of the AI SDK, including core functionalities and specific provider integrations like @ai-sdk/alibaba and @ai-sdk/amazon-bedrock. The updates also impact internal harness packages such as @ai-sdk/harness-pi@1.0.105, @ai-sdk/harness-cline@1.0.30, and @ai-sdk/harness-codex@1.0.105.
Why It Matters
These updates enhance the robustness and flexibility of the Vercel AI SDK, offering developers more control over model usage in batch processes and improving the reliability of image result handling. The enforcement of tool choices in streaming text can lead to more predictable and consistent AI interactions.
What To Do
- Review the ai@7.0.94 release notes for detailed implementation specifics regarding per-request models in batch.
- Test the streamText function to observe the enforced tool choices.
- Note the updated dependency versions for @ai-sdk/provider, @ai-sdk/gateway, @ai-sdk/provider-utils, and @ai-sdk/alibaba.
- Observe how the new retry mechanisms handle empty image results and terminal prompt blocks from Google and Google Vertex.