[Issue]: How to Store Multiple Autogen Agents and Histories in Azure for Multi-User Environment

[chenna-nefer]

Describe the issue

I created an agentic workflow using Autogen, where multiple agents are stored in global variables and used as required. This worked well locally, but now I'm moving to a multi-user environment and setting up authentication for multiple users. I'm facing challenges on how to store these agents and their histories in Azure for each user.

What I’ve Tried:

Blob Storage: Successfully stored chat history in .pkl files in Azure Blob Storage.
MongoDB: Attempted to use MongoDB for storing chat histories and agent configurations but struggled to store the agents themselves.

How can I efficiently manage the storage and retrieval of these multiple autogen agents and chat histories in Azure to ensure scalability and persistent storage?
Are there best practices or existing solutions for storing multiple agents and their configurations in a database like MongoDB or any other recommended storage solutions?

Steps to reproduce

No response

Screenshots and logs

No response

Additional Information

No response

[giannisha]

Hi @chenna-nefer. This doesn't seem so an Autogen issue. Can you give some context on how you are running your agentic workflow on Azure? Have you packaged it in a web api (e.g. Flask) or sth like Prompflow ?

[chenna-nefer]

Hi @giannisha

I’m running the agentic workflow on Azure packaged as a Flask web API.

Additional Context:

Setup: Multiple agents managed and stored in global variables, now transitioning to a multi-user environment.
Storage Attempts:
Blob Storage: Successfully storing chat histories in .pkl files.
MongoDB: Facing issues with storing agents due to unserializable registered functions.
Challenges:

Efficiently storing and retrieving agents and chat histories for multiple users.
Ensuring scalability and persistent storage.
Questions:

Best practices for storing multiple agents and their configurations in a database like MongoDB or other solutions?
Should I consider dynamically creating agents each time a user interacts with them, storing only the message history and state?

[giannisha]

@chenna-nefer thanks for the extra info! What do you want to accomplish by keeping the same agent instance?

This seems to me a more logical approach. You can consider also dependency-injector to better handle singleton and transient instances in your Flask api.

Should I consider dynamically creating agents each time a user interacts with them, storing only the message history and state?

[Priya-Issar]

Hi @chenna-nefer , @giannisha
I’m also working on a use case involving text-to-SQL query generation using Autogen agents with a sequential flow. My setup includes 5-6 agents(assistant and user proxy agents).

I’m currently exploring how to store short-term memory for a specific chat session to maintain context and enhance the interaction experience. Could you explain how you’re managing memory in your setup? Are you using any particular storage method ?

If you’ve implemented something similar, it would be great if you could share the resources or sample code for the same!
Looking forward to your insights!

[giannisha]

Hello @Priya-Issar,

The simplest approach might be to store your chat messages in a Redis (or similar) cache manually after each interaction and load them before a new one.
The following two examples refer to Zep and Mem0 products which offer something more integrated and can give you some idea how to approach it.

agent memory with zep
agentchat memory with mem0

I hope it helps.

[Priya-Issar]

Hi @giannisha,

I’ve checked the articles you mentioned, but I’m still unclear about what type of chat history is most useful to save for a Text-to-SQL use case.

Here’s the sequential agent flow I’m using in detail:

1. User Proxy Agent: Takes the user question from the frontend.
2. Assistant: Understands the use case and prepares a task plan to solve the query.
3. Assistant: Retrieves the DDL from the vector database based on the user query(text question of user received through user-proxy
   agent).
4. Assistant: Uses the DDL retrieved in the previous step to generate the SQL query.

My primary aim is to store the chat history for the duration of one user-chat session. One approach I’m considering is saving the user’s question and the generated SQL query. However, I’m not sure how this would enhance the active chat session experience.

Could you suggest the best way to integrate memory storage at each step or agent in such a workflow? or any ideas about , If there’s a better method to maintain memory for a chat session.

Looking forward to your suggestions!