Google expanded document inputs and multimodal function responses in the Gemini API on November 5, 2026. The update gives agents more ways to receive evidence from files and tools, with processing rules that differ by format.
Documents do not all become the same input
The update adds PPT, ODP, and ODS input support, alongside page-image processing for DOC files. Google’s document guide distinguishes rendered pages from spreadsheet data. In particular, the ODS path supplies data and formulas from the first sheet. Existing processing for other supported Office formats is unchanged.
That distinction affects what an application can ask reliably. A rendered page preserves visual arrangement, while extracted spreadsheet values serve a different purpose. Accepting a file does not establish that every hidden sheet, embedded object, or external dependency has become available to the model.
Tool results can carry richer evidence
The November notice also expands function results to include audio, video, Office, and OpenDocument content. Supported file references let a tool return a Files API URI rather than always embedding bytes. The GenerateContent and Interactions interfaces expose their own documented content structures.
These are input and transport changes. They do not automatically make an application’s connector fetch the correct file or give it permission to read a user’s storage. The surrounding application still owns access, retrieval, and the association between a result and its source.
Test the evidence the model actually receives
Our implementation analysis is to build a small fixture collection from the documents a product expects to handle. Include a slide deck with important visual relationships, a document with tables, and a spreadsheet where a conclusion depends on more than one sheet. Check which information reaches the model before judging the answer.
Ask questions with known, independently checked answers. Then include cases where the supported processing omits something essential. A useful system should identify that limitation and request another input, rather than inventing content for the missing portion.
For tool-generated files, preserve the original file identity and the version used for each result. A user can replace a document while an agent is working. Without a clear association, a correct analysis of the earlier version may be presented as a finding about the current one.
Keep file access bounded
Apply the same ownership checks, size limits, and retention decisions used for other uploads. Verify that a returned file reference remains accessible only within its intended scope and that failures remain visible. A convenient URI should not become an unrestricted route to unrelated files.
The release broadens what developers can build with document-aware agents. Its practical value comes from matching each format’s processing behavior to a specific task and retaining enough provenance to check the resulting answer.