Update app.py

app.py

@@ -3,7 +3,7 @@ import gradio as gr
 import requests
 import inspect
 import pandas as pd
-from smolagents import CodeAgent, DuckDuckGoSearchTool, InferenceClientModel # Make sure InferenceClientModel is imported
+from smolagents import CodeAgent, DuckDuckGoSearchTool, InferenceClientModel
 
 # --- Constants ---
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
@@ -24,50 +24,40 @@ class BasicAgent:
             # use_structured_outputs_internally=True # Optional: Can sometimes improve performance for capable models, but might also cause issues if the model doesn't strictly adhere. Test this if needed.
         )
 
-        # --- REFINED SYSTEM PROMPT ---
-        SYSTEM_PROMPT = """You are a highly capable and precise AI assistant. Your primary goal is to answer questions accurately and concisely, following strict formatting rules.
+        # --- REVISED SYSTEM PROMPT FOR EXACT MATCH ---
+        # This prompt is designed to make the agent output ONLY the answer,
+        # without any preambles, thoughts, or "FINAL ANSWER:" tags.
+        # This is critical for exact match scoring.
+        SYSTEM_PROMPT = """You are an ultra-concise and accurate AI assistant for the GAIA benchmark.
+        Your sole purpose is to provide the exact correct answer to the given question, and NOTHING ELSE.
+        Do not include any introductory phrases, thoughts, explanations, or "FINAL ANSWER:" tags.
+        Your output must be the answer only.
 
-        **Here's how you should approach each question:**
-        1.  **Understand the Request:** Read the question carefully to identify what information is needed and what type of answer is expected (number, string, or list).
+        **Here's how you should determine and format your answer:**
+
+        1.  **Understand the Question:** Carefully read the question to identify the precise information needed.
         2.  **Gather Information:**
-            * First, try to answer based on your internal knowledge if the information is common or involves simple logic/arithmetic.
-            * If the question requires factual knowledge, current information, or is complex, **you MUST use the `DuckDuckGoSearchTool`**. Formulate precise search queries to get the most relevant results.
-            * Analyze the search results meticulously to extract the exact information required.
-        3.  **Reason Step-by-Step:** Before providing the final answer, explicitly state your thought process. This helps in complex reasoning. Use a "Thought:" section.
-        4.  **Formulate FINAL ANSWER:** Construct your final answer strictly according to these rules:
-            * **Template:** Always end your entire response with `FINAL ANSWER: [YOUR FINAL ANSWER]`.
-            * **Numbers:** If the answer is a number, write it in digits ONLY, without commas, unit symbols (e.g., "$", "%", "km", "kg"), or text (e.g., "one" instead of "1"). For example, "12345" not "12,345" or "12345 dollars".
-            * **Strings (words):** If the answer is a string, use as few words as possible. Do NOT use articles (a, an, the). Do NOT use abbreviations. If a number is part of a string, write it in plain English words (e.g., "twenty" instead of "20"), unless it's a specific numerical value (like a year or an ID).
-            * **Comma-separated lists:** If the answer is a list, apply the above rules to each element in the list and separate elements with a comma and a single space (e.g., "apple, banana, orange" or "one, two, three").
-
-        **Examples of desired FINAL ANSWER formats:**
-
-        * **Question:** What is the capital of Japan?
-            * **Thought:** I know the capital of Japan is a specific city.
-            * **FINAL ANSWER: Tokyo**
-
-        * **Question:** How many days are in a common year?
-            * **Thought:** A common year has a fixed number of days.
-            * **FINAL ANSWER: three hundred sixty five**
-
-        * **Question:** Who developed Python and in what year?
-            * **Thought:** This requires factual recall of an inventor and a year. I will use DuckDuckGoSearchTool if uncertain.
-            * **FINAL ANSWER: Guido van Rossum, nineteen ninety one**
-
-        * **Question:** List three primary colors.
-            * **Thought:** I need to list common primary colors.
-            * **FINAL ANSWER: red, yellow, blue**
-
-        * **Question:** Calculate 15 multiplied by 7.
-            * **Thought:** This is a simple arithmetic calculation.
-            * **FINAL ANSWER: one hundred five**
-
-        Your detailed thought process (after "Thought:") should lead directly to the accurate `FINAL ANSWER` based on the specified format.
+            * Use your internal knowledge if the answer is straightforward or involves simple arithmetic.
+            * **Crucially, use the `DuckDuckGoSearchTool`** if the question requires external knowledge, current facts, or complex research. Formulate the most effective search queries to find exact information.
+            * Analyze search results meticulously to extract the precise fact(s).
+        3.  **Formulate the Answer (Strict Formatting Rules):**
+            * **Numbers:** If the answer is a number, provide it in digits ONLY. Do NOT use commas for thousands separators, currency symbols (e.g., "$"), percentage signs (e.g., "%"), or any other units (e.g., "kg", "meters").
+                * *Example 1 (number):* If the question asks "What is the square root of 81?", your answer should be "9".
+                * *Example 2 (large number):* If the question asks "What is the population of City X (approx)?", and you find "1,234,567", your answer must be "1234567".
+            * **Strings (words):** If the answer is text, use the fewest words possible. Do NOT include articles (a, an, the). Do NOT use abbreviations. If a number is part of a string (e.g., "twenty books"), write the digit in plain English words (e.g., "twenty").
+                * *Example 1 (string):* If the question asks "Who invented the telephone?", your answer should be "Alexander Graham Bell".
+                * *Example 2 (string with number in text):* If the question asks "How many primary colors are there?", your answer should be "three".
+            * **Comma-separated lists:** If the answer is a list of items, apply the above rules to each item, and separate them with a comma and a single space (ee.g., "red, green, blue").
+                * *Example:* If the question asks "List two types of fruit.", your answer should be "apple, banana".
+
+        Your final output *must* be only the answer, following these exact rules. No surrounding text.
         """
         self.agent.prompt_templates["system_prompt"] = self.agent.prompt_templates["system_prompt"] + SYSTEM_PROMPT
             
     def __call__(self, question: str) -> str:
         print(f"Agent received question (first 50 chars): {question[:50]}...")
+        # The key change: The agent's run method should directly produce the final answer,
+        # without additional wrapping, due to the hyper-specific prompt.
         final_answer = self.agent.run(question)
         print(f"Agent returning final answer: {final_answer}")
         return final_answer
@@ -99,8 +89,9 @@ def run_and_submit_all( profile: gr.OAuthProfile | None):
 
     # In the case of an app running as a hugging Face space, this link points toward your codebase ( usefull for others so please keep it public)
     # IMPORTANT: Update this to YOUR SPACE_ID if you've cloned it!
-    agent_code = f"https://huggingface.co/spaces/{os.getenv('SPACE_ID', 'ghanemfaouri/Final_Assignment_Template')}/tree/main"
-    print(f"Agent code will link to: {agent_code}") # Added a print for clarity
+    # Ensure this link is correct for your public space for verification.
+    agent_code = f"https://huggingface.co/spaces/{os.getenv('SPACE_ID', 'YOUR_HF_USERNAME/YOUR_SPACE_NAME')}/tree/main"
+    print(f"Agent code will link to: {agent_code}")
 
     # 2. Fetch Questions
     print(f"Fetching questions from: {questions_url}")
@@ -135,6 +126,16 @@ def run_and_submit_all( profile: gr.OAuthProfile | None):
             continue
         try:
             submitted_answer = agent(question_text)
+            # IMPORTANT: Post-process the submitted_answer if the model still adds unwanted text.
+            # While the prompt aims to prevent it, sometimes models are stubborn.
+            # This stripping makes sure only the raw answer is sent.
+            submitted_answer = submitted_answer.strip()
+            # Remove any trailing "FINAL ANSWER:" or "FINAL ANSWER:" if model still generates it.
+            if submitted_answer.startswith("FINAL ANSWER:"):
+                submitted_answer = submitted_answer[len("FINAL ANSWER:"):].strip()
+            # You might need more sophisticated stripping depending on how the model misbehaves
+            # For example, if it adds thoughts, you'd need to extract the last line or a specific pattern.
+
             answers_payload.append({"task_id": task_id, "submitted_answer": submitted_answer})
             results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": submitted_answer})
         except Exception as e: