add script to batch inference for study; hard-coded weather call

ai_agent_converter.py

@@ -0,0 +1,74 @@
+import json
+
+from azure.ai.projects.models import (RunStepType, MessageRole, ThreadMessage, MessageTextContent,
+                                      MessageTextDetails, RunStepFunctionToolCall, RunStepFunctionToolCallDetails, OpenAIPageableListOfRunStep,
+                                      RunStep, RunStepMessageCreationDetails, RunStepMessageCreationReference, RunStepCompletionUsage, RunStepToolCallDetails,
+                                      OpenAIPageableListOfThreadMessage)
+
+class ThreadMessageEncoder(json.JSONEncoder):
+    def default(self, obj):
+        if isinstance(obj, (RunStepFunctionToolCallDetails, OpenAIPageableListOfRunStep, RunStep,
+                            RunStepMessageCreationDetails, RunStepMessageCreationReference, RunStepCompletionUsage,
+                            RunStepToolCallDetails, OpenAIPageableListOfThreadMessage)):
+            return obj.__dict__["_data"]
+        if isinstance(obj, RunStepFunctionToolCall):
+            return obj.__dict__["_data"]
+        if isinstance(obj, MessageTextDetails):
+            return obj.__dict__["_data"]
+        if isinstance(obj, MessageTextContent):
+            return obj.__dict__["_data"]
+        if isinstance(obj, ThreadMessage):
+            json_data = obj.__dict__["_data"]
+            if obj.__dict__.get("tool_calls"):
+                json_data["tool_calls"] = obj.__dict__["tool_calls"]
+            return json_data  # or implement a method to convert to a dictionary
+        return super().default(obj)
+
+# project_client.telemetry.enable(destination=sys.stdout)
+
+class AIAgentConverter:
+    def __init__(self, project_client):
+        self.project_client = project_client
+
+    def convert(self, thread_id, filter_run_id=None):
+        """
+        Fetches all messages in a thread and converts them to JSON.
+        if filter_run_id is provided, only messages from that run are included. Assuming all messages before the last assistant messages for that run are part of that run.
+        """
+        messages = self.project_client.agents.list_messages(thread_id=thread_id)
+        with open("messages.json", 'w') as file:
+            json.dump(messages, file, indent=4, cls=ThreadMessageEncoder)
+
+        messages = messages.data
+
+        assistant_message_index_for_run = None
+        for i in range(0, len(messages)):
+            message = messages[i]
+            print(f"Message: {message.content}")
+            message_id = message.id
+            message_type = message.role
+            run_id = message.run_id
+            if message_type == MessageRole.AGENT:
+                if filter_run_id is not None and run_id == filter_run_id:
+                    assistant_message_index_for_run = i
+                tool_calls = []
+                if filter_run_id is None or run_id == filter_run_id:
+                    run_details = self.project_client.agents.list_run_steps(thread_id=thread_id, run_id=run_id)
+                    with open("run_details.json", 'w') as file:
+                        json.dump(run_details, file, indent=4, cls=ThreadMessageEncoder)
+                    for run_step in run_details.data:
+                        print(f"Run step: {run_step.type}")
+                        if run_step.type == RunStepType.MESSAGE_CREATION:
+                            print(f"Assistant message: {run_step.step_details.message_creation.message_id}")
+                        elif run_step.type == RunStepType.TOOL_CALLS:
+                            tool_calls.extend(run_step.step_details.tool_calls)
+                            print(f"Tool call: {run_step.step_details.tool_calls}")
+                    message.tool_calls = tool_calls
+
+        evaluation_data = messages[assistant_message_index_for_run:] if assistant_message_index_for_run is not None else messages,
+        json_data = json.dumps(
+            messages[assistant_message_index_for_run:] if assistant_message_index_for_run is not None else messages,
+            cls=ThreadMessageEncoder)
+        with open("proposed_evaluation_data.json", 'w') as file:
+            json.dump(evaluation_data, file, indent=4, cls=ThreadMessageEncoder)
+        return json.loads(json_data)