When near end of turns, force LLM to wrap it up with final answer guidelines

Author: pseudotensor

openai_server/autogen_2agent_backend.py

@@ -120,6 +120,7 @@ def run_autogen_2agent(query=None,
         autogen_code_restrictions_level,
         agent_venv_dir,
         temp_dir,
+        max_turns=autogen_max_turns,
         agent_tools_usage_hard_limits=agent_tools_usage_hard_limits,
         agent_tools_usage_soft_limits=agent_tools_usage_soft_limits,
     )
@@ -167,6 +168,7 @@ def code_writer_terminate_func(msg):
                               human_input_mode="NEVER",
                               is_termination_msg=code_writer_terminate_func,
                               max_consecutive_auto_reply=autogen_max_consecutive_auto_reply,
+                              max_turns=autogen_max_turns,
                               )
 
     code_writer_agent = H2OConversableAgent("code_writer_agent", **code_writer_kwargs)
@@ -177,7 +179,7 @@ def code_writer_terminate_func(msg):
         # setup planning agents
         code_writer_kwargs_planning = code_writer_kwargs.copy()
         # terminate immediately
-        update_dict = dict(max_consecutive_auto_reply=1)
+        update_dict = dict(max_consecutive_auto_reply=1, max_turns=None)
         # is_termination_msg=lambda x: True
         code_writer_kwargs_planning.update(update_dict)
         code_writer_agent = H2OConversableAgent("code_writer_agent", **code_writer_kwargs_planning)

openai_server/autogen_utils.py

@@ -669,7 +669,10 @@ def __init__(
             default_auto_reply: Union[str, Dict] = "",
             description: Optional[str] = None,
             chat_messages: Optional[Dict[Agent, List[Dict]]] = None,
+            max_turns: Optional[int] = None,
     ):
+        self.max_turns = max_turns
+        self.turns = 0
         code_execution_config = (
             code_execution_config.copy() if hasattr(code_execution_config, "copy") else code_execution_config
         )
@@ -831,6 +834,7 @@ def _generate_code_execution_reply_using_executor(
             iostream = IOStream.get_default()
             iostream.print(delta)
             output += delta
+        self.turns += 1
         return valid, output
 
     def __generate_code_execution_reply_using_executor(
@@ -887,6 +891,10 @@ def __generate_code_execution_reply_using_executor(
                     # force immediate termination regardless of what LLM generates
                     self._is_termination_msg = lambda x: True
                     return True, self.final_answer_guidelines()
+            if self.max_turns is not None and self.turns >= self.max_turns - 1:
+                # one before final allowed turn, force LLM to stop
+                self._is_termination_msg = lambda x: True
+                return True, self.final_answer_guidelines()
 
             num_code_blocks = len(code_blocks)
             if num_code_blocks == 1:

src/version.py

@@ -1 +1 @@
-__version__ = "5ea362c96da92eff13e1c2252e8030515d71c085"
+__version__ = "9349803e8bfd5b3d42ab1e5a2bfadcb649774891"