From the source
Google DeepMind has widened what tools can hand back to Gemini models, letting function responses carry Office and OpenDocument files, audio and video alongside the images, PDFs and plain text already supported.
- In the GenerateContent API, function response parts now also accept fileData, so a tool can return a Files API URI instead of inline bytes.
- The same parts gained displayName on both inlineData and fileData, and video parts accept a mediaProcessing setting of STATIC or AGENTIC.
- The Interactions API was updated in parallel: function_result steps now take document, audio and video content blocks in addition to text and image, delivered either as inline data or a URI.
- Video blocks there are available in v1beta and v1alpha and accept a processing setting.




