what is the prompt management feature for ADK?

[rajib76]

When we use AWS Strands agents, we can use AWS Bedrock Prompt management to manage the prompts outside the agent code. What is the equivalent service/feature in ADK or GCP?

[juszzz]

@rajib76 have you seen https://docs.cloud.google.com/vertex-ai/generative-ai/docs/model-reference/prompt-classes?

[rajib76]

Thanks for sharing this. I see this is under vertex AI package. Do we then need to use both vertex AI and google-genai package. I was also looking for creating the prompts, compare them through google console instead of programmatic way. This package(vertexai) also looks like is not working

[marcopellegrino-devoteam]

My guess is that Agent Engine will have a prompt management feature one day, the Vertex AI solutions doesn't work well.
You could implement LangFuse for now.

[hangfei]

Move to discussion.

Here are some features: https://google.github.io/adk-docs/context/ and more are coming!

[adk-bot]

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

TLDR: ADK follows a "code-first" philosophy, so there isn't a direct, GUI-based equivalent to AWS Bedrock's prompt management. Instead, prompts are managed directly within your Python code, offering flexibility and version control.

---

Hello! That's a great question.

The Agent Development Kit (ADK) for Python is designed with a "code-first" approach. This means that configurations, including agent instructions and prompts, are defined directly in Python code. This allows you to version, test, and manage your prompts just like any other part of your application's source code. [4]

In a typical ADK agent, the prompts are part of the Agent object's instruction parameter. You can see this in many of the ADK examples. [2]

For more dynamic scenarios where you might need to modify prompts before they are sent to the model, ADK provides a powerful callback system. Specifically, the before_model_callback allows you to inspect and modify the request that will be sent to the LLM, which includes the prompt. [7] This is useful for:

• Injecting dynamic information into the prompt.
• Implementing safety guardrails.
• Adapting prompts based on the conversational state.

Additionally, you can use ADK Plugins to implement cross-cutting concerns like request/response modification. [14, 11]

While ADK integrates with Google Cloud services like Vertex AI for model backends [13], grounding [6], and long-term memory [22], the core prompt definition and management remain within the code-centric framework of ADK itself. The primary way to manage prompts is by treating them as code, which has the benefits of fitting into standard DevOps workflows.