Set agent seed for deterministic output across runs

[brianverbanckinect]

Hey guys

I was wondering if anyone knew how to set up the seed for LLM agents. I have been experimenting with following code but can't seem to get the right setting nor find it online.
agent_kwargs = {
"name": agent_name,
"model": model,
"description": description,
"instruction": instruction,
"tools": tools or [],
"output_schema": output_schema,
"generate_content_config": types.GenerateContentConfig(
seed=0
)
}

 Let me know if anyone has had the same issue!

[adk-bot]

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

TLDR: The seed parameter is not supported for achieving deterministic outputs in the way you are attempting. To get deterministic outputs from the LLM, you should set temperature=0 in the generation_config.

---

Hello! I see you're looking to get deterministic outputs from your LLM agent.

While LlmAgent is generally non-deterministic due to the nature of the underlying language model, you can make its responses deterministic for a given model version by setting the temperature to 0.

The generate_content_config or generation_config parameter should be used to pass a dictionary or a GenerationConfig object with the temperature set. The seed parameter is not a recognized option for this configuration.

Here is how you can adjust your code:

from google.generativeai import types

agent_kwargs = {
    "name": agent_name,
    "model": model,
    "description": description,
    "instruction": instruction,
    "tools": tools or [],
    "output_schema": output_schema,
    "generation_config": types.GenerationConfig(
        temperature=0.0
    )
}

Or, more simply, using a dictionary:

agent_kwargs = {
    "name": agent_name,
    "model": model,
    "description": description,
    "instruction": instruction,
    "tools": tools or [],
    "output_schema": output_schema,
    "generation_config": {"temperature": 0.0}
}

Please note that while this will make the output deterministic for the specific model version you are using, the output is not guaranteed to be the same across different model versions.

Let me know if you have any other questions!

[brianverbanckinect]

Actually looking for the same response across multiple runs of the model/agents. So a low temperature does not seem like the fix in my use-case