Adding autogen integration#57
Conversation
|
@manfredcalvo I know it's not specific to your PR necessarily because you used the openaisdk as a template, but the pattern of having agents/ Feels very unclear. Do you have thoughts on how it could be restructured? |
|
|
||
| FC_AGENT_DEFAULT_YAML_CONFIG_FILE_NAME = "function_calling_agent_config.yaml" | ||
|
|
||
| class FunctionCallingAgent(mlflow.pyfunc.PythonModel): |
There was a problem hiding this comment.
Do we need to wrap this in a pyfunc? I thought MLFlow supported Autogen
There was a problem hiding this comment.
MLflow seems to support only auto-log for autogen; check that here: https://mlflow.org/docs/latest/_modules/mlflow/autogen.html. The tricky part is that to create an agent with two actors (assistant and user), you need two conversable agents, so your model ends up having two objects instead of one, like in langchain/langgraph. I want to explore group chats in the future and see if I can add both actors and have only one object that wraps both.
For sure we could change that and keep the config inside the each folder (agents, data_pipeline). Like: agents/config/FunctionCallingAgentConfig data_pipeline/config/DataPipelineConfig, so we have info config local to each component (agent, data pipeline, etc). I did not change that structure because I believe we should agree on a specific approach, make the necessary changes, and then move it to a shared folder that can be accessed by different libraries. The data pipeline folder is identical in both the langgraph and openai_sdk, so once we refactor that we can change the structure. |
2 participants
Related Issues/PRs
None
What changes are proposed in this pull request?
Implement function calling agent using Autogen. Tested with UC tool and vector search retriever tool
How is this PR tested?
I conducted manual tests using the review app and the playground in the Databricks workspace.