Working with Images, Audio, and Video in Gemini for the Agentic Architect Exam

GCP Study Hub
November 14, 2026

Gemini on Agent Platform accepts text, images, audio, and video in the same request, which allows the model to reason across different media types directly. Designing reliable multimodal requests for the Professional Agentic Architect exam comes down to keeping related inputs together when they must be compared, passing large files by reference through Cloud Storage instead of embedding bytes inline, and enabling audio_timestamp when audio-only transcripts need accurate timecodes.

Reasoning across images and text in one request

Because Gemini accepts images and text in the same request, one call can reason across a specific image and the text written about it. Sending the image, the related document, and the instructions together lets the model compare them directly and point out statements in the text that the image does not support. When the files live in Cloud Storage, referencing them by URI also keeps large images clear of inline request size limits.

The general principle is that the model can only check what it actually receives. A comparison holds up when both inputs reach the same model in the same request. Any design that hands the model a stand-in for one of the inputs, such as a text description of the image or only the extracted text of a document, limits the check to whatever that stand-in happens to capture. The same applies when one input goes to a vision model and the other to a separate text model, since neither model ever sees both.

It also helps to keep comparison separate from search. Agent Retrieval is built for finding similar items across a large collection, which is the right tool when the matching item is unknown. When the specific file and its corresponding text are already known, the work is a direct comparison that Gemini handles in a single request, with no training data or tuned model involved.

Passing large media by Cloud Storage URI

Embedding media file bytes inline makes the request body large. Inline data uses base64 encoding, which makes the data about a third larger than the original file, so long videos or full-length recordings can push the payload past the allowed size and fail with an HTTP 413 payload too large error. Large media should be passed by reference instead.

Storing each file in Cloud Storage and passing its gs:// URI together with its MIME type as file data lets Gemini on Agent Platform read the media directly from Cloud Storage rather than from the request body. Every request stays small, which makes it possible to process long videos and recordings that no inline request could hold, and the model still receives the whole file, so it can reason across a video's audio and frames together. For Gemini to access the file, the Cloud Storage bucket must be in the same project as the request or be publicly readable.

A 413 error is about the size of the request body, so the useful test for any change is whether it takes the bytes out of that body. Keeping the media inline, however it is encoded, compressed, or divided within the request, leaves long recordings exposed to the same limit. Client and generation settings do not move the limit either. A request timeout governs how long the client waits, and maxOutputTokens limits the length of the model's response rather than the size of the input.

Getting accurate timestamps from audio-only files

Google lists a specific limitation for audio-only files: accurate timestamps require setting audio_timestamp to true in the generation config of each request. Without this parameter enabled, timestamps in audio-only transcriptions can be inaccurate. Because the setting lives in the generation config, it has to be part of every request that needs timecodes.

With audio_timestamp enabled, Gemini can return a transcript in a timecode, speaker, and caption format. Audio recordings can stay in Cloud Storage and be passed by gs:// URI with an audio MIME type, the same reference pattern used for large video, and a single prompt can include one audio file of up to about 8.4 hours of audio.

The setting is specific to audio-only input, and it is a documented requirement rather than a quality adjustment. Other generation settings serve different purposes. A lower temperature, for example, makes output more repeatable, but repeatable output is not the same as correct output.

Applying these patterns on the exam

Each media pattern rests on its own principle. Inputs that must be compared belong together in a single Gemini request so the model sees all of them at once. Large media travels by reference, as gs:// URIs with MIME types from a same-project or publicly readable Cloud Storage bucket, so request bodies stay small. Accurate timecodes for audio-only files, up to about 8.4 hours per prompt, come from turning on audio_timestamp in the generation config.

Our Professional Agentic Architect practice exams test working with images, audio, and video in Gemini alongside Gemini generation configuration and Agent Retrieval, with detailed explanations that drill these distinctions.

Get tips and updates from GCP Study Hub

arrow