Gemini API Adds Office Files to Tool Results: Formats and Limits

Google's November 5 Gemini API release notes add three document inputs: legacy PowerPoint PPT, OpenDocument Presentation ODP and OpenDocument Spreadsheet ODS. Word DOC files also gain page-image processing. The changes are available now, according to the announcement.

The same update expands what tools can return to a model. Function responses can carry Office and OpenDocument files, audio and video. GenerateContent also gains file-URI support for these responses, so a tool can return a Files API reference instead of embedding the file's bytes in the request.

For teams building document assistants, this changes the retrieval step. A tool that finds a presentation or recording can hand the file back for inspection without first reducing it to a text summary. The application still needs to know what the model will actually receive.

File support does not mean identical processing

Google's document-processing guide describes four different paths:

• DOC, PPT and ODP: Every page or slide is rendered as an image, allowing the model to inspect layout, charts and pictures across the file.
• DOCX, DOTX, ODT and PPTX: The model receives text from the whole file, with embedded images also included for DOCX and PPTX. Pages and slides are not rendered, so layout and some visual content may be lost.
• XLSX and XLS: Only the first sheet is read. It becomes CSV containing cell values, without formulas or formatting.
• ODS: Only the first sheet is read, but its cell data and formulas are included.

Processing for DOCX, DOTX, ODT, PPTX, XLSX and XLS has not changed in this release. The same processing rules apply whether a person supplies the file directly or a tool returns it. Existing file-size and page limits still apply to the whole file.

A spreadsheet assistant should therefore identify the requested sheet before passing the workbook along. Google recommends moving that sheet first or saving it separately. An answer based on a summary tab should not be presented as an analysis of every worksheet.

Match the response to your API

The GenerateContent function-calling guide places media inside the parts of a functionResponse. A part contains either inlineData or fileData. Both can carry displayName; references from the structured response require a unique name, referenced only once.

Video parts can set mediaProcessing to STATIC or AGENTIC. When using agentic video processing, preserve the full returned response in subsequent history, including its media-processing tool parts.

The Interactions API version uses a different structure: content blocks inside the result of a function_result step. The new block types are document, audio and video, with inline base64 data or a uri. Video blocks are supported in v1beta and v1alpha, and use a processing field.

The guides describe multimodal function responses for Gemini 3 series models. Check the selected model and API version rather than assuming that ordinary function-calling support covers every media feature. Unsupported MIME types return 400 INVALID_ARGUMENT.

Treat file references as temporary

The Files API documentation lists a 2 GB per-file storage maximum and 20 GB per project. Uploaded files expire after 48 hours. Those storage limits do not override format-specific processing limits. Files API storage is free; model inference is separate.

A durable job queue needs a way to refresh an expired reference from the original source. Keep the source identifier alongside the URI, and check permissions again before retrieving private material. Saving a URI alone is not an archival strategy.

Before switching a production pipeline, build a small test set around the actual failure modes: a chart-heavy deck, a workbook whose useful sheet is second, a formula-dependent ODS file and a long recording returned by a tool. Compare the answers with known facts and measure latency, token use and errors.

The update can remove conversion work, but that is a hypothesis to test on your files. These release notes do not establish the accuracy or cost of your document workflow.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn AI infrastructure and developer-tool updates into clear technical writing and product stories.

Book a 30-minute call.