Transform to add an agent's name into the message content

[marklysze]

The recent addition of non-OpenAI client classes has expanded the utility of AutoGen to a lot of different models through other cloud providers as well as the ability to run it with local LLMs.

However, the name field on each message is removed as only OpenAI can accept and utilise that field.

This means that functionality, like the GroupChat's select speaker, isn't as effective as the LLM may not be able to determine which 'agent' belongs to each message.

This PR introduces a transform that can add an agent's name to the start or end of message's text content. As this is a transform, it isn't permanent and doesn't change the original message.

How it works:

from autogen.agentchat.contrib.capabilities import transform_messages, transforms

# Create a transform
name_transform = TextMessageContentName(position="start", format_string="'{name}' said:\n")

# Create the TransformMessages
context_handling = transform_messages.TransformMessages(
            transforms=[
                name_transform
            ]
        )

# Add it to an agent so when they run inference it will apply to the messages
context_handling.add_to_agent(my_agent)

# Example
# Messages before
[
        {"role": "system", "content": "I am the system."},
        {"role": "user", "name": "charlie", "content": "I think the sky is blue."},
        {"role": "user", "name": "mary", "content": "The sky is red."},
        {"role": "user", "name": "bob", "content": "The sky is crimson."},
    ]

# Messages going into LLM (where name is removed, non-OpenAI clients)
[
        {"role": "system", "content": "I am the system."},
        {"role": "user", "content": "'charlie' said:\nI think the sky is blue."},
        {"role": "user", "content": "'mary' said:\nThe sky is red."},
        {"role": "user", "content": "'bob' said:\nThe sky is crimson."},
    ]

Documentation to be done.

Dependency

For the GroupChat select speaker to be able to utilise this, PR microsoft#2719 needs to be merged first.

Why are these changes needed?

To provide the necessary name context for non-OpenAI inference.

Related issue number

Checks

• I've included any doc changes needed for https://microsoft.github.io/autogen/. See https://microsoft.github.io/autogen/docs/Contribute#documentation to build and test documentation locally.
• I've added tests (if relevant) corresponding to the changes introduced in this PR.
• I've made sure all auto checks have passed.

(3334)

[sonichi]

Thanks. Could you include the code snippet in "how it works" in docstr and a test?

[Hk669]

can we just combine these checks.

Suggested change

	assert isinstance(position, str) and position is not None
	assert position in ["start", "end"]
	assert isinstance(position, str) and position in ["start", "end"]

[marklysze]

I'll just check that None is handled correctly.

[marklysze]

Okay, None handled okay in both cases, thanks - will change

[Hk669]

Suggested change

	assert isinstance(format_string, str) and format_string is not None
	assert "{name}" in format_string
	assert isinstance(format_string, str) and "{name}" in format_string

[Hk669]

@marklysze , ive also noticed, gpt-4o-mini doesnt support name and breaks the flow.

[marklysze]

@marklysze , ive also noticed, gpt-4o-mini doesnt support name and breaks the flow.

This was noted in another issue, do you have a code sample? I couldn't replicate an issue with name included.

[marklysze]

Thanks. Could you include the code snippet in "how it works" in docstr and a test?

Thanks @sonichi, I've added the code snipped to the class docstring.

For the tests, the test_transforms.py has a number of tests for including the name related to position and filtering. Was there some you thought were missing? Reference

[marklysze]

A note: use of this transform is included in website doc changes in PR #2