Implementing a new role in Taskweaver as part of ext_role

[Rmaram07]

Hi, I am trying to implement a new role ("Reviewer agent") that will check user query, plan, generated output to see if they align well, if not communicate to planner to make necessary changes. I am trying to implement .reply() method and seems like the reply method is not being executed when I run the application. I followed documentation here and here. I made necessary changes to all the script files.

When I tried to run the recepta role or echo role that is already present, I don't see recepta at the planner/code interpretor role level, so my another question is, are these roles output part of planner output or we see a separate entity when I run the application? Can you share more detailed implementation of any particular role? I am having hard time debugging the taskweaver code base.

Looks like code in taskweaver/session/session.py starting from line 183 is not understanding how to call the reply() of Reviewer agent that i started to build.

if recipient == "Planner": reply_post = self.planner.reply( self.memory, prompt_log_path=os.path.join( self.workspace, f"planner_prompt_log_{chat_round.id}_{post.id}.json", ), ) elif recipient in self.worker_instances.keys(): reply_post = self.worker_instances[recipient].reply( self.memory, prompt_log_path=os.path.join( self.workspace, f"code_generator_prompt_log_{chat_round.id}_{post.id}.json", ), ) else: raise Exception(f"Unknown recipient {recipient}")

Below is the reviewer agent python script code that I had written. Can anyone guide me on implementing an ext_role as a new agent that can communicate with planner to make necessary changes before output is shown to the user? If you look at the image below, I have put self.logger.info("reviewer.reply() method called"), I don't see this in my logs when I run the application and start the conversation session, that means my reply method is not being executed.

[liqul]

When I tried to run the recepta role or echo role that is already present, I don't see recepta at the planner/code interpretor role level, so my another question is, are these roles output part of planner output or we see a separate entity when I run the application?

Hi @Rmaram07, I didn't get this question clearly. I'm not sure what do you mean by saying 'run the recepta role or echo role'. If you have read the docs in your link, you know that the planner can talk with other roles in posts, and the output of a role (e.g., echo) will not be part of the planner's output, only part of the conversation of the planner.

When implementing a new role, the typical way of using it is to have the planner talking to this role when necessary (based on the prompt of this role), not directly calling it without involving the planner. But you can actually do this by configuring only the new role name in the session configuration.

[Rmaram07]

Hi liqul, I followed the documentation for adding a new role and ask it to talk to planner. In case of code interpreter, if planner decides to generate code, it calls code interpreter and gets the result. Just like that for reviewer role I've implemented, it needs to communicate to planner and I'm having hard time triggering it. I mean, the reviewer agent code is not getting executed. I have also configured the role in session configuration.

[Rmaram07]

Thanks Liqul for the clarity. I was able to implement a new role sucessfully but I would like to know, if we can default a role to run every time before sending the output to the user? Is there a way we can do that? just like when we ask any code related question, then only code interpreter will be triggered. in this case, can we always call one role (reviewer) every time by default before sending answer to the user?

[liqul]

Thanks Liqul for the clarity. I was able to implement a new role sucessfully but I would like to know, if we can default a role to run every time before sending the output to the user? Is there a way we can do that? just like when we ask any code related question, then only code interpreter will be triggered. in this case, can we always call one role (reviewer) every time by default before sending answer to the user?

@Rmaram07 TaskWeaver is not designed to be a generic multi-agent framework where flexible orchestration is supported. That means you can not control the topology of the roles or agents. So, in short, there is no 100% reliable way of running a role before sending the response to the user. We had a similar scenario before where we need to summarize the entire conversation at the end of a chat. What we did is to instruct the planner (in its prompt) that it must call this role to do the summarize before responding to the user and it works in most of the time, but as I said, not guarranteed.

[Rmaram07]

okay got it. Could you please answer these questions below?

1. Where is the shared memory actually stored (locally in the planner container?) can it be stored in a shared location?
2. How are the planner and code interpreter interacting with shared memory? Do both read and write? what exactly are they reading/writing from the shared memory?
3. what content is in shared memory, is it the entire conversation history with intermediate steps? Is there additional conversation history somewhere else? Is it just the most recent chat question? Is it just the most recent planner step?

[liqul]

okay got it. Could you please answer these questions below?

1. Where is the shared memory actually stored (locally in the planner container?) can it be stored in a shared location?
2. How are the planner and code interpreter interacting with shared memory? Do both read and write? what exactly are they reading/writing from the shared memory?
3. what content is in shared memory, is it the entire conversation history with intermediate steps? Is there additional conversation history somewhere else? Is it just the most recent chat question? Is it just the most recent planner step?

I'm not going to explain all the details here on the shared memory. It is stored in the memory so not persisted anywhere. It is designed to be used by the roles so that there is another channel, besides the back-and-forth posts, to exchange necessary information.

If you want to understand how it is used in TaskWeaver, I encourage you to run it and debug locally to see when it get triggered and how information is organized.

[Rmaram07]

Hi @liqul, Can you help me with session management of task weaver outside the memory like ADLS? I know its currently implemented as "in-memory" and there's no support for external data sources. Could you guide me implementing a custom solution?

[liqul]

ADLS

I didn't get your question on managing session memory. Could you explain in a more detailed way?

[Rmaram07]

I would like to place all the sessions files in the cloud i.e Azure data lake storage.