Several breaking changes have been introduced to the GenAI schemas: frequency_penalty, presence_penalty, and cached_content have been removed from GenaiVertexV1beta1GenerationConfig and GenaiVertexV1beta1ModelInteraction. Developers relying on these for fine-tuning output diversity or context caching will need to adjust their implementations. On the feature side, a new publishers.v1.responses method allows for forwarding arbitrary HTTP requests to deployed models, supporting both streaming and non-streaming cases.
Additionally, the retrieval framework for agents has been expanded with GenaiVertexV1beta1RetrievalCallDelta and RetrievalCallStep, which now support new tools including EXA_AI_SEARCH and PARALLEL_AI_SEARCH. Video generation capabilities are also being refined with new experimental configurations for spatialAlignment and colorAlignment. Finally, inferenceEventLoggingConfig has been updated to clarify that it now covers all publisher-model traffic, including both Provisioned Throughput and on-demand usage.
