Input settings in public spec, not announced by Google

Unannounced settings that give fine-grained control over spreadsheet, document and audio inputs have been found in the public specification of Google's "Gemini API" for developers. @aitrackerbot, a bot that detects API changes, reported them on X on Nov. 16 (early Nov. 17 Japan time). According to the post, there are settings to choose the sheet and cell range to read, read formulas or a rendered image of the sheet, set page ranges and clip audio.

The Gemini API's schema just got controls for spreadsheet and document inputs: pick a sheet and cell range, read formulas or a rendered image of the sheet, set page ranges, and clip audio. The docs still say only an XLSX's first sheet is read, as CSV.

ORIENS confirmed on Nov. 17 three settings, "SpreadsheetMetadata," "DocumentMetadata" and "AudioMetadata," in the "Discovery document," in which Google publishes the structure of the API for machines. All of this is unconfirmed, leak-stage information: Google has not announced it, and the changelog does not mention the additions. ORIENS has not run tests with an API key, and it is not known whether specifying the fields actually works.

Fields to choose sheet, cell range, formulas or image

According to the Discovery document, SpreadsheetMetadata is a field that configures "static" preprocessing of an input spreadsheet. It can be placed both in the input part of a request (Part) and in the part that returns the result of a function call (FunctionResponsePart).

sheetIndex specifies which sheet to read, counted from 1 in workbook order, and visibleCells specifies a cell range in A1 notation, such as "A1:D10." If no cell range is specified, the full occupied area of the sheet is used, according to the description.

contentLayers selects what content to return. The three options are "DATA" for cell values, merges and formats; "FORMULAS" for formula expressions and their computed results; and "VISUAL" for a rendered image of the sheet. If left empty, DATA and FORMULAS are the defaults.

Page ranges for documents, segments for audio

DocumentMetadata configures preprocessing of documents, with startPage and endPage specifying the pages to read, counted from 1. Both end pages are included in the range.

contentLayers selects either "PAGE_IMAGE" for rendered page images or "PAGE_TEXT" for extracted page text. If left empty, PAGE_IMAGE is the default, and without this setting the default processing applies (page images for PDFs), according to the description.

AudioMetadata configures preprocessing of audio, with startOffset and endOffset specifying start and end positions measured from the start of the audio stream. The "clip audio" in the bot's post corresponds to this field. It is separate from the existing videoMetadata for video.

Official documentation still says "first sheet only"

Google's official documentation, "Document understanding," still says that for Excel XLSX and XLS files, only the first sheet is read. The contents are converted to CSV (comma-separated text data), so the model receives cell values but not formulas or formatting. Other sheets are ignored, and the page advises moving a different sheet to the first position or saving it as a separate file to analyze it. It does not mention the new settings.

The bot's post also noted that the documentation still says only the first sheet is read, as CSV. If the fields in the public specification are provided as described, developers would be able to specify the range to read and whether to include formulas without reordering sheets or splitting files. For corporate developers who have AI read spreadsheet files, this is a factor in deciding whether to revisit their current procedures.

Google's changelog runs through Nov. 12 and contains no entry announcing these settings. On Nov. 5, Google expanded acceptance of files such as PowerPoint 97-2003 (PPT) and OpenDocument formats, but said at the time that processing for XLSX and XLS was "unchanged." On the same day, it enabled documents, audio and video to be returned to the model as function results. The documents do not say how the new settings relate to these changes.

Documents differ on handling when no sheet is specified

Documents describe differently what happens when no sheet number is specified. The Discovery document (v1beta and v1alpha) says that if it is not set, all sheets are extracted. Meanwhile, Google's Python and JavaScript SDKs (development toolkits) define the same type on their main branches and say that if it is not set, the first sheet is used. ORIENS has not tested the behavior, and which is the actual behavior is unknown.

Not in stable v1; supported models and pricing unknown

The three settings appear in the beta v1beta and alpha v1alpha versions of the Discovery document but not in the stable v1. All three versions carry the revision "20261116," but this value identifies the document's current revision, and the date the fields were added is unknown. The bot wrote "just got," but when the fields were added could not be confirmed.

None of the documents state the supported models, pricing or release timing. It is also unknown how the input volume or cost would change when choosing the sheet image (VISUAL) or page images. On Nov. 7, ORIENS also reported another unannounced field, "proactiveVideo," in the Gemini API's public specification, but whether it is part of the same change as these fields is unknown.

The entries in each document are summarized below. The Discovery document and the SDKs define the three settings, but the official documentation and changelog do not mention them.

Whether the three input settings are mentioned (checked by ORIENS in public documents as of Nov. 17)
Public documentMentions the three settingsDescription when no sheet is specified
Discovery document (v1beta)YesAll sheets are extracted
Discovery document (v1alpha)YesAll sheets are extracted
Discovery document (v1)No―
Python SDK (main branch)YesThe first sheet is used
JavaScript SDK (main branch)YesThe first sheet is used
Document understandingNoOnly the first sheet of XLSX and XLS is read, as CSV
Gemini API changelogNo―