functionality of manual history cleaning by user proxy added

[Grigorij-Dudnik]

What is this and how to use it?

Adds ability to clear agents' message history by user proxy manually writing "clear history" within message of user feedback.

Works for a groupchat when one of the agents is user proxy, recommended human_input_mode="ALWAYS".

When user asked to provide manual feedback, he can write "clear history" in his message to remove all the messages from agents history. Writing "clear history <agent_name>" will clear the history of messages only for selected agent. Writing "clear history <nr_of messages_to_preserve>" clears all the history except of last <nr_of messages_to_preserve> messages. Writing "clear history <agent_name> <nr_of messages_to_preserve>" will clear the history for selected agent except last <nr_of messages_to_preserve> messages.

User can noramally write other things in his message. "clear history" string and parameters will cutted out from user reply before passing it to the chat.

Why are these changes needed?

History of messages tends to grow endless very fast during group conversation. According to my experience, many of old messages became useless and just provide noise. Providing unnecesary history to LLM is bad practice because of:

• Model attention distraction. Valuable information could be lost in the mess of noise. Model performance could significantly drop.
• Processing unnecesary history causes huge costs. According to my experience, I'm spending 20-30$ on 3-4h session with Autogen using gpt-4-turbo.
• Also causes long response times.
• There is a risk of context window overflow, especially for models with small/medium context window lenghthes.

That's why having possibility to manage history during conversation is important.

Related issue number

It could be solution or partial solution for #156, but not only.

Checks

• I've included any doc changes needed for https://microsoft.github.io/autogen/. See https://microsoft.github.io/autogen/docs/Contribute#documentation to build and test documentation locally.
• I've added tests (if relevant) corresponding to the changes introduced in this PR.
• I've made sure all auto checks have passed.

[codecov-commenter]

Codecov Report

Attention: 5 lines in your changes are missing coverage. Please review.

Comparison is base (1ab2354) 32.48% compared to head (fe813da) 45.25%.

Files	Patch %	Lines
autogen/agentchat/groupchat.py	92.10%	1 Missing and 2 partials ⚠️
autogen/agentchat/conversable_agent.py	77.77%	1 Missing and 1 partial ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##             main    #1230       +/-   ##
===========================================
+ Coverage   32.48%   45.25%   +12.76%     
===========================================
  Files          41       41               
  Lines        4907     4950       +43     
  Branches     1120     1201       +81     
===========================================
+ Hits         1594     2240      +646     
+ Misses       3187     2521      -666     
- Partials      126      189       +63     

Flag	Coverage Δ
unittests	45.19% <89.36%> (+12.74%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[Grigorij-Dudnik]

@microsoft-github-policy-service agree

[Grigorij-Dudnik]

@sonichi what do you think now? Everything is improved, I hope I answered all questions. Can we merge?

[sonichi]

@sonichi what do you think now? Everything is improved, I hope I answered all questions. Can we merge?

Thanks. Running final tests.

[Grigorij-Dudnik]

Seems like a problems with web browser. Have no idea why, changes are totally unrelated. I have no expertise in web browsing functionality, please help.

EDIT: I see it is a known bug with WebSurfer.

[afourney]

Yeah there is a bug in of websurfer's tests. The fix is here, but can't be merged because of another bug in an unrelated notebook code formatting check:

#1417