A workbook can upload successfully while leaving most of its sheets outside the model’s input.
Google expanded Gemini API document support on November 5, adding PPT, ODP, and ODS files. The API changelog also records page-image processing for DOC and more file types in tool results. For developers, the useful change is the set of formats a workflow can accept. The important limit is how each format reaches the model.
A file extension now determines more than acceptance
The document-processing documentation separates page rendering from text extraction. DOC, PPT and ODP files become images of their pages or slides. DOCX, DOTX, ODT and PPTX follow a text-processing path instead. Google also documents embedded-image extraction for DOCX and PPTX.
That difference changes the evidence available to answer the question. A rendered slide can expose its visual arrangement. Extracted text can preserve words without preserving a chart or the relationship between objects. Accepting both files does not make their representations equivalent.
| Files and scope | Model input |
|---|---|
| DOC, PPT, ODP Whole document | Images of pages or slides |
| DOCX, DOTX, ODT, PPTX Whole document | Text; images from DOCX and PPTX |
| XLSX, XLS First sheet | CSV cell values; no formulas or styles |
| ODS First sheet | Cell data and formulas |
The first sheet becomes an application decision
For a workbook with an assumptions sheet followed by quarterly results, the documented spreadsheet path supplies the assumptions sheet alone. An answer about the quarterly results could therefore lack the relevant cells even though the file was accepted.
A practical ingestion policy should record which sheet contains the requested evidence. If the task spans sheets, the application needs a separate extraction or conversion step that includes those sheets. Keep the original file available so a reviewer can trace the answer back to the correct cells.
Formulas need their own check. ODS supplies formulas from its first sheet; XLSX and XLS supply values without formulas. A question about how a total was calculated needs different evidence from a question asking for that total.
Tool responses can carry documents too
Google’s function-calling documentation extends multimodal tool results to Office and OpenDocument files, audio and video. In GenerateContent, function responses can use file references as an alternative to inline bytes. The document-processing rules also apply when a supported document arrives through a function response.
This lets a retrieval tool return an actual file for the model to examine. It also makes the application responsible for selecting the right representation. Returning a workbook from a tool does not remove its first-sheet boundary.
Choose the representation before trusting the answer
For document automation, I would start with the evidence the task needs: readable text, page appearance, cell values, or calculation logic. Then choose a supported path that supplies it. We reviewed the documentation, but we haven’t run a cross-format accuracy test.
Which file causes the most trouble in your workflow: a scanned-looking presentation, a modern slide deck, or a workbook with several sheets? Share the format and the task in the comments.




