ConversableAgent "send" function "context" question

[bonadio]

Hi
I was looking at the "send" function in ConversableAgent,

conversable_agent

But I did not understand the example with the lambda function

{ "content": lambda context: context["use_tool_msg"], "context": { "use_tool_msg": "Use tool X if they are relevant." } }

If I run this code

user_proxy.send( { "content": lambda context: context["use_tool_msg"], "context": { "use_tool_msg": "Use tool X if they are relevant." }}, order_assistant, request_reply=True)

I get

`user_proxy (to order_assistant):

Use tool X if they are relevant.

`

What is the use case for this type of message?

Thanks

[sonichi]

@JieyuZ2 would you like to explain this use case?

[JieyuZ2]

The example is to show how to use the context mechanism. In general, the context mechanism allows you to change any user message in the conversation history. For example, the context could be

def content_function(use_b):
    if use_b:
        return str_a + str_b
    else:
        return str_a  

Then, the first message you send to the model could be

{ "content": content_function, "context": { "use_b": True } }

At certain turn of the conversation, you want to switch the first message to str_a only, you can send

{ "content": "current message", "context": { "use_b": False } }

Then, when calling the LLM, the first message would be changed to str_a only. (note: when using chat LLMs, every time we have to send it the whole conversation history.)

One use case of this mechanism is, the initial message contains some useful but long information (str_b), and you only want to send this information to the LLM once. After the LLM's first reply, you can turn off the str_b by sending a new context with your current message to the LLM to save the cost and token used.

In my EcoAssistant project (https://github.com/JieyuZ2/EcoAssistant/), I append some demonstrations to the first message sent to the LLM, then in the subsequent conversation / calls of LLM, I turn off the demonstration in the first message to save the number of inputting tokens.

Let me know if you have any question!

[bonadio]

Hi @JieyuZ2, your note that every time we send the whole conversation, was what I was missing.
The code in your project is really good. You should post at the Discord channel in the Tutorials.
Thanks

[zuoyifan132]

@JieyuZ2 I wonder if send and receive function can be used in an instance of autogen.GroupChat? Since the chat is managed by the group chat manager, is there a way to intervene the group chat?

[JieyuZ2]

I think this new society_of_mind agent may fit your need
https://github.com/microsoft/autogen/blob/main/notebook/agentchat_society_of_mind.ipynb

[zuoyifan132]

Thanks for the solution and that perfectly solve my issue. One concern about the autogen framework: does the framework deal with the long context(long group chat history since all agents share with same history) problem?

[JieyuZ2]

this may help https://github.com/microsoft/autogen/blob/main/notebook/agentchat_compression.ipynb

[sonichi]

@gagb has made a new way to handle long context for any agent using a new capability.
cc @kevin666aa

[zuoyifan132]

@JieyuZ2 does the outside world chat history all be visible from SocietyOfMind internal monologue group?