Fix: switch to create_tool_calling_agent + AgentExecutor

Replaces LangGraph create_react_agent which caused malformed tool calls
on Groq. LangChain's tool-calling agent uses proper JSON tool format.
Also adds handle_parsing_errors=True and enforces non-empty answers.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app.py

@@ -3,38 +3,39 @@ import gradio as gr
 import requests
 import pandas as pd
 from langchain_groq import ChatGroq
-from langgraph.prebuilt import create_react_agent
+from langchain.agents import create_tool_calling_agent, AgentExecutor
 from langchain_community.tools import DuckDuckGoSearchRun, WikipediaQueryRun
 from langchain_community.utilities import WikipediaAPIWrapper
 from langchain_experimental.tools import PythonREPLTool
 from langchain.tools import tool
-from langchain_core.messages import SystemMessage, HumanMessage
+from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
 
 # --- Constants ---
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 
 SYSTEM_PROMPT = """You are a general AI assistant tasked with answering questions accurately.
 
-Reason step by step, then finish your answer with exactly this format:
+Use tools to research and reason step by step. When you have a final answer, output ONLY:
 FINAL ANSWER: [your answer]
 
 Rules for FINAL ANSWER:
 - Numbers: no commas, no units unless specified
 - Strings: no articles (a/an/the), no abbreviations, write digits in plain text
 - Lists: comma-separated, apply above rules per element
-- Be precise and concise — the answer is evaluated by exact match"""
+- Be precise — evaluated by exact match
+- Never say "I cannot answer" — always attempt an answer"""
 
 
 @tool
 def fetch_task_file(task_id: str) -> str:
-    """Fetch a file associated with a GAIA task by its task_id. Use this when the question references an attachment, file, or additional data."""
+    """Fetch a file associated with a GAIA task by its task_id. Use when the question references an attachment, file, image, or additional data."""
     try:
         response = requests.get(f"{DEFAULT_API_URL}/files/{task_id}", timeout=15)
         if response.status_code == 200:
             content_type = response.headers.get("content-type", "")
-            if any(t in content_type for t in ["text", "json", "csv", "xml"]):
+            if any(t in content_type for t in ["text", "json", "csv", "xml", "python"]):
                 return response.text[:8000]
-            return f"Binary file ({content_type}) — cannot read as text"
+            return f"Binary file detected ({content_type}) — cannot read as text"
         return "No file found for this task"
     except Exception as e:
         return f"Error fetching file: {e}"
@@ -49,17 +50,26 @@ class BasicAgent:
             PythonREPLTool(),
             fetch_task_file,
         ]
-        self.agent = create_react_agent(llm, tools)
-        print("BasicAgent initialized with LangGraph + Groq (llama-3.3-70b-versatile).")
+        prompt = ChatPromptTemplate.from_messages([
+            ("system", SYSTEM_PROMPT),
+            ("human", "{input}"),
+            MessagesPlaceholder("agent_scratchpad"),
+        ])
+        agent = create_tool_calling_agent(llm, tools, prompt)
+        self.executor = AgentExecutor(
+            agent=agent,
+            tools=tools,
+            max_iterations=15,
+            handle_parsing_errors=True,
+            verbose=True,
+        )
+        print("BasicAgent initialized with LangChain tool-calling agent + Groq (llama-3.3-70b-versatile).")
 
     def __call__(self, question: str, task_id: str = "") -> str:
         full_question = f"[Task ID: {task_id}]\n\n{question}" if task_id else question
         print(f"Running agent on task {task_id}: {question[:80]}...")
-        result = self.agent.invoke(
-            {"messages": [SystemMessage(content=SYSTEM_PROMPT), HumanMessage(content=full_question)]},
-            {"recursion_limit": 50},
-        )
-        raw_answer = result["messages"][-1].content
+        result = self.executor.invoke({"input": full_question})
+        raw_answer = result.get("output", "")
         if "FINAL ANSWER:" in raw_answer:
             answer = raw_answer.split("FINAL ANSWER:")[-1].strip()
         else:
@@ -184,7 +194,7 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
 
 # --- Gradio Interface ---
 with gr.Blocks() as demo:
-    gr.Markdown("# Agent Evaluation Runner — LangGraph + Groq")
+    gr.Markdown("# Agent Evaluation Runner — LangChain + Groq")
     gr.Markdown(
         """
         **Instructions:**
@@ -192,7 +202,7 @@ with gr.Blocks() as demo:
         1. Log in to your Hugging Face account using the button below.
         2. Click 'Run Evaluation & Submit All Answers' to fetch questions, run the agent, and submit.
 
-        **Agent:** LangGraph ReAct agent with Groq (llama-3.3-70b-versatile)
+        **Agent:** LangChain tool-calling agent with Groq (llama-3.3-70b-versatile)
         **Tools:** DuckDuckGo search, Wikipedia, Python REPL, File fetcher
 
         ---