checkpoint: Function calling program fails. Needs investigation.

src/coder_agent.py

@@ -23,8 +23,10 @@
 from llama_index.core.tools import ToolSelection, ToolOutput
 from llama_index.core.workflow import Event
 
+from pydantic import BaseModel
 from typing_extensions import TypedDict
 
+from llama_index.core.program import FunctionCallingProgram
 
 from code_executor import CodeExecutor
 
@@ -41,6 +43,12 @@ class TestCase(TypedDict):
     outputs: str
 
 
+class CoderAgentOutput(BaseModel):
+    reasoning: str
+    pseudocode: str
+    code: str
+
+
 class InputEvent(Event):
     input: list[ChatMessage]
     test_cases: list[TestCase] = None
@@ -68,6 +76,15 @@ def __init__(
 
         self.tools = [FunctionTool.from_defaults(self.evaluate)]
 
+        self.pydantic_code_generator = FunctionCallingProgram.from_defaults(
+            output_cls=CoderAgentOutput,
+            llm=self.llm,
+            verbose=True,
+            prompt_template_str="{input}",
+            tool_choice=self.tools[0],
+            allow_parallel_tool_calls=True,
+        )
+
         self.memory = ChatMemoryBuffer.from_defaults(llm=llm)
         self.sources = []
 
@@ -90,6 +107,9 @@ async def handle_llm_input(self, ev: InputEvent) -> ToolCallEvent | StopEvent:
         ic(ev)
         chat_history = ev.input
 
+        ic(chat_history[-1].content)
+        pydantic_call = self.pydantic_code_generator(input=chat_history[-1].content)
+        ic(pydantic_call)
         response = await self.llm.achat_with_tools(
             self.tools, chat_history=chat_history
         )