deep research team in autogenstudio fails with error "This model's maximum context length is 128000 tokens. However, your messages resulted in 209101 tokens"

[dctmfoo]

What happened?

i sent a query to the deep research team in autogenstudio using "Test Team". After a couple of mins of researching it failed with the below error

raise self._make_status_error_from_response(err.response) from None
openai.BadRequestError: Error code: 400 - {'error': {'message': "This model's maximum context length is 128000 tokens. However, your messages resulted in 209101 tokens. Please reduce the length of the messages.", 'type': 'invalid_request_error', 'param': 'messages', 'code': 'context_length_exceeded'}}

What did you expect to happen?

I expected the research to complete and print the results.

How can we reproduce it (as minimally and precisely as possible)?

In Test Team of Deep Research Team, enter this query and press enter

"are the valuations of Nifty Small Cap favorable to buy now?"

AutoGen version

latest

Which package was this bug in

AutoGen Studio

Model used

gpt-4o

Python version

No response

Operating system

No response

Any additional info you think would be helpful for fixing this bug

No response

[victordibia]

Hi @dctmfoo

Good observations. A good starting point is to review how the deep research team works

Specifically, the research_assistant agent in the team uses several tools - google_search and fetch_webpage. These tools in turn return the content of a webpage as markdown . All of this data is added to context. This is what leads to hitting the maximum model context limit error you see as many turns are taken -

There are two potential fixes.

• modify tools to limit the amount of text they return. This could be as simple as setting the max length in the tools parameter e.g, only return the first 1000 characters. I could be more involved .. e.g, using an LLM to extract the most important bit.
• updating the model_context input parameter for the research_agent. Specifically, you can pass in a BufferedChatCompletionContext and configure it to use only the last n messages while making an LLM call. This reduces the amount of context that is sent an LLM each time that agent makes an LLM call

Would you be interested in exploring any one of these and sharing your findings?

[victordibia]

@dctmfoo let me know if any of the proposed ideas above work for you