Multimodal Image use in ADK

[Avishkarp23]

I want to know how do we use Image as an input to an agent in a sequential workflow agent as I already have an agent previously working

[adk-bot]

Response from ADK Answering Agent (experimental, answer may be inaccurate)

TLDR: To use an image as input in a sequential workflow, you should use ADK's Artifact management system to handle the image file. The first agent can save the image as an Artifact and pass its reference to the next agent through the shared session state.

---

Hello! I'm a bot that can help answer questions about ADK. Based on the ADK documentation, here’s how you can handle image inputs in a sequential workflow:

ADK has built-in support for multimodal inputs, including images. [1] The recommended way to manage binary data like images is by using the ADK Artifacts feature. Artifacts are designed to handle file-like binary data that you need to persist, version, or share between different parts of your application. [2]

In a SequentialAgent, agents are executed in a fixed order, and they can share information using the session state. An earlier agent can write its output to the session state, and a subsequent agent can read that information from the context state. [3]

To pass an image from one agent to the next in your sequence, you can follow these steps:

1. Ingest the Image as an Artifact: The first agent in your sequence (or a tool it uses) should take the image input and create an Artifact from it. This is the standard way to handle user-uploaded files or any binary data within ADK. [2]

2. Save the Artifact to Session State: Once the Artifact is created, the agent should save a reference to this Artifact into the shared session state. This is typically done by specifying an output_key for the agent's result.

3. Access the Artifact in the Next Agent: The next agent in the SequentialAgent workflow can then access the Artifact reference from its context state. With the Artifact reference, it can retrieve the image data and use it for its processing tasks.

By using Artifacts and the shared session state, you can effectively pass images and other binary data between agents in your sequential workflow.

I hope this helps you build your multimodal agent!

---

Citations:
[1] https://google.github.io/adk-docs/get-started/about/
[2] https://google.github.io/adk-docs/artifacts/
[3] https://google.github.io/adk-docs/agents/multi-agents/