Update app.py

app.py

@@ -9,104 +9,280 @@ from smolagents import CodeAgent, DuckDuckGoSearchTool, InferenceClientModel, to
 # --- Constants ---
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 
+
 # ==========================================
-# 🚀 TOOL 1: RICERCA E LETTURA WEB AVANZATA
+# 🔧 TOOL 1: LETTURA WEBPAGE
 # ==========================================
 @tool
 def visit_webpage(url: str) -> str:
-    """Visits a webpage and extracts its main clean text. Use this to read Wikipedia, news, or articles.
+    """Visits a webpage and extracts its main clean text content.
+    Use this to read Wikipedia pages, news articles, or any online resource.
     Args:
-        url: The URL of the webpage to visit.
+        url: The full URL of the webpage to visit (e.g. 'https://en.wikipedia.org/wiki/...')
     """
     try:
-        headers = {'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/91.0.4472.124 Safari/537.36'}
+        headers = {
+            'User-Agent': (
+                'Mozilla/5.0 (Windows NT 10.0; Win64; x64) '
+                'AppleWebKit/537.36 (KHTML, like Gecko) '
+                'Chrome/91.0.4472.124 Safari/537.36'
+            )
+        }
         response = requests.get(url, headers=headers, timeout=15)
         response.raise_for_status()
-        
         soup = BeautifulSoup(response.text, 'html.parser')
-        # Rimuove tutto ciò che non è testo utile
         for element in soup(["script", "style", "nav", "footer", "header", "aside"]):
             element.extract()
-            
         text = soup.get_text(separator='\n', strip=True)
-        # Prende i primi 15000 caratteri di puro testo informativo
-        return text[:15000] 
+        return text[:15000]
     except Exception as e:
         return f"Error reading the webpage: {str(e)}"
 
+
+# ==========================================
+# 🎬 TOOL 2: TRASCRIZIONE YOUTUBE
+# ==========================================
+@tool
+def get_youtube_transcript(video_url: str) -> str:
+    """Fetches the transcript/captions of a YouTube video.
+    Use this whenever the question refers to a YouTube video URL.
+    Args:
+        video_url: The full YouTube video URL (e.g. 'https://www.youtube.com/watch?v=...')
+    """
+    try:
+        from youtube_transcript_api import YouTubeTranscriptApi
+        match = re.search(r'(?:v=|youtu\.be/)([^&\n?#]+)', video_url)
+        if not match:
+            return "Could not extract video ID from URL."
+        video_id = match.group(1)
+        transcript_list = YouTubeTranscriptApi.get_transcript(video_id, languages=['en', 'it', 'auto'])
+        full_text = ' '.join([entry['text'] for entry in transcript_list])
+        return full_text[:10000]
+    except Exception as e:
+        return f"Transcript not available: {str(e)}"
+
+
+# ==========================================
+# 📂 TOOL 3: DOWNLOAD FILE DA GAIA
+# ==========================================
+@tool
+def download_task_file(task_id: str) -> str:
+    """Downloads and reads the file attached to a GAIA task (if any).
+    Returns the text content of the file or a description of it.
+    Args:
+        task_id: The task_id string from the GAIA question.
+    """
+    try:
+        file_url = f"https://agents-course-unit4-scoring.hf.space/files/{task_id}"
+        response = requests.get(file_url, timeout=15)
+        if response.status_code == 404:
+            return "No file attached to this task."
+        response.raise_for_status()
+        content_type = response.headers.get('Content-Type', '')
+
+        if 'text' in content_type or 'json' in content_type or 'csv' in content_type:
+            return response.text[:10000]
+
+        if 'pdf' in content_type:
+            try:
+                import io
+                import PyPDF2
+                pdf_reader = PyPDF2.PdfReader(io.BytesIO(response.content))
+                text = ''
+                for page in pdf_reader.pages:
+                    text += page.extract_text() or ''
+                return text[:10000]
+            except Exception:
+                return f"PDF downloaded ({len(response.content)} bytes) but could not extract text."
+
+        if 'image' in content_type:
+            return f"Image file attached (content-type: {content_type}). Size: {len(response.content)} bytes. Cannot parse directly."
+
+        # fallback: try as plain text
+        try:
+            return response.content.decode('utf-8')[:10000]
+        except Exception:
+            return f"Binary file attached (content-type: {content_type}, size: {len(response.content)} bytes)."
+    except Exception as e:
+        return f"Error downloading task file: {str(e)}"
+
+
+# ==========================================
+# 🔍 PRE-PROCESSING DOMANDA
+# ==========================================
+def preprocess_question(question: str) -> str:
+    """Handles special question formats before sending to the agent."""
+
+    # 1. Testo scritto al contrario (reversed text)
+    stripped = question.strip()
+    reversed_q = stripped[::-1].strip()
+    if any(word in reversed_q.lower() for word in ['answer', 'write', 'what', 'who', 'how', 'find', 'list', 'if you']):
+        if len(reversed_q) > 10:
+            print(f"[PRE-PROCESS] Testo invertito rilevato. Versione corretta: {reversed_q[:80]}")
+            return reversed_q
+
+    return question
+
+
 # ==========================================
 # 🧠 IL SUPER AGENTE
 # ==========================================
 class SuperAgent:
     def __init__(self):
         print("Inizializzazione del SUPER Agente AI in corso...")
-        
-        # 1. Modello
-        self.model = InferenceClientModel(model_id="Qwen/Qwen2.5-Coder-32B-Instruct")
-        
-        # 2. Tools
-        self.tools = [DuckDuckGoSearchTool(), visit_webpage]
-        
-        # 3. Agente (Potenza massima, importazioni analitiche sbloccate)
+
+        # Modello principale — Llama 3.3 70B per ragionamento general-purpose
+        self.model = InferenceClientModel(
+            model_id="meta-llama/Llama-3.3-70B-Instruct"
+        )
+
+        # Modello di fallback leggero per risposte dirette senza tools
+        self.fallback_model = InferenceClientModel(
+            model_id="Qwen/Qwen2.5-72B-Instruct"
+        )
+
+        # Tools disponibili
+        self.tools = [
+            DuckDuckGoSearchTool(),
+            visit_webpage,
+            get_youtube_transcript,
+            download_task_file,
+        ]
+
+        # Agente principale
         self.agent = CodeAgent(
             tools=self.tools,
             model=self.model,
-            max_steps=20,  # Aumentati i tentativi a 20!
+            max_steps=10,
             additional_authorized_imports=[
-                "requests", "bs4", "json", "time", "math", "datetime", 
-                "pandas", "numpy", "re", "csv", "urllib"
+                "requests", "bs4", "json", "time", "math", "datetime",
+                "pandas", "numpy", "re", "csv", "urllib", "collections",
+                "itertools", "string", "unicodedata"
             ]
         )
-        
-        # 4. Prompt Estremo per GAIA
-        self.prompt_template = """
-        You are an elite AI data analyst solving the GAIA benchmark. 
-        You are provided with a question. You MUST use your tools to find the answer.
-        
-        CRITICAL RULES FOR YOUR FINAL OUTPUT:
-        1. EXACT MATCH ONLY: Output ONLY the final answer. Nothing else.
-        2. If the answer is a number, return JUST the number (e.g., '14' or '1998').
-        3. If the answer is a name/word, return JUST the word.
-        4. NEVER use phrases like "The answer is", "Based on my search", or "FINAL ANSWER:".
-        5. If the question requires math, date calculation, or text processing, write the Python code to solve it internally.
-        
-        Question: {question}
-        """
-
-    def __call__(self, question: str) -> str:
-        print(f"\n[DOMANDA RICEVUTA]: {question[:80]}...")
+
+        # Prompt ottimizzato per GAIA
+        self.prompt_template = """You are an expert AI assistant solving the GAIA benchmark evaluation.
+Your goal is to find the EXACT correct answer to the question below.
+
+STRATEGY:
+- If the question references a YouTube video URL → use get_youtube_transcript tool first.
+- If the question references a website or Wikipedia → use visit_webpage tool.
+- If the question seems to have an attached file → use download_task_file with the task_id.
+- For factual questions → use DuckDuckGoSearchTool to search, then visit_webpage to confirm.
+- For math, date arithmetic, text manipulation → write Python code to compute the answer directly.
+- If the text looks reversed or scrambled → reverse it first with Python.
+
+OUTPUT RULES (CRITICAL):
+1. Return ONLY the final answer. No explanation, no preamble.
+2. Numbers: return just the number (e.g. '3' or '1998').
+3. Names/words: return just the word or name.
+4. Lists: return comma-separated values.
+5. NEVER say "The answer is", "FINAL ANSWER:", "Based on", etc.
+
+Question: {question}
+"""
+
+        # Prompt diretto per fallback (senza tools)
+        self.direct_prompt = """You are an expert assistant. Answer the following question with ONLY the final answer.
+No explanation. No preamble. Just the answer itself.
+
+If the text is reversed, reverse it and answer accordingly.
+If it is a math question, compute and give only the result.
+If it is a factual question, give only the fact.
+
+Question: {question}
+Answer:"""
+
+    def _clean_answer(self, raw: str) -> str:
+        """Rimuove prefissi verbosi e formattazione indesiderata dalla risposta."""
+        answer = str(raw).strip()
+
+        # Rimuovi prefissi verbosi comuni
+        prefixes = [
+            "the answer is", "final answer:", "answer:", "final answer is",
+            "the requested word is", "the highest number is", "the result is",
+            "based on", "according to", "the word is", "the name is",
+            "the number is", "the correct answer is", "the response is",
+        ]
+        lower = answer.lower()
+        for prefix in prefixes:
+            if lower.startswith(prefix):
+                answer = answer[len(prefix):].strip()
+                lower = answer.lower()
+                break
+            # Cerca anche nel mezzo della stringa come ultima occorrenza
+            idx = lower.rfind(prefix)
+            if idx != -1:
+                candidate = answer[idx + len(prefix):].strip()
+                # Solo se il candidate è breve (vera risposta finale)
+                if len(candidate) < 200:
+                    answer = candidate
+                    lower = answer.lower()
+                    break
+
+        # Toglie punto finale
+        if answer.endswith('.') and not re.search(r'\d\.$', answer):
+            answer = answer[:-1].strip()
+
+        # Toglie markdown e virgolette extra
+        answer = answer.replace("**", "").strip('"').strip("'").strip()
+
+        return answer
+
+    def _direct_answer(self, question: str) -> str:
+        """Chiede direttamente al modello senza usare l'agente con tools."""
+        try:
+            prompt = self.direct_prompt.format(question=question)
+            response = self.fallback_model([{"role": "user", "content": prompt}])
+            # InferenceClientModel restituisce un ChatMessage
+            if hasattr(response, 'content'):
+                raw = response.content
+            else:
+                raw = str(response)
+            return self._clean_answer(raw)
+        except Exception as e:
+            print(f"[FALLBACK MODEL ERROR]: {e}")
+            return "I don't know"
+
+    def __call__(self, question: str, task_id: str = "") -> str:
+        print(f"\n[DOMANDA]: {question[:100]}...")
+
+        # Pre-processing
+        processed_question = preprocess_question(question)
+
+        # Se task_id disponibile, aggiungilo al contesto del prompt
+        context = ""
+        if task_id:
+            context = f"\nNote: This question has task_id '{task_id}'. Use download_task_file('{task_id}') if a file might be attached.\n"
+
+        full_question = processed_question + context
+
+        # --- TENTATIVO 1: Agente completo con tools ---
         try:
-            formatted_prompt = self.prompt_template.format(question=question)
+            formatted_prompt = self.prompt_template.format(question=full_question)
             raw_answer = self.agent.run(formatted_prompt)
-            final_answer = str(raw_answer).strip()
-            
-            # --- FILTRO ANTI-BLABLA ---
-            # Se l'LLM si ostina a inserire testo, lo forziamo a tacere tagliando le frasi comuni.
-            prefixes_to_cut = ["The answer is", "FINAL ANSWER:", "Answer:", "final answer is", "The requested word is", "The highest number is"]
-            for prefix in prefixes_to_cut:
-                if prefix.lower() in final_answer.lower():
-                    idx = final_answer.lower().rfind(prefix.lower()) + len(prefix)
-                    final_answer = final_answer[idx:].strip()
-                    
-            # Toglie il punto finale se l'ha messo per sbaglio (es. "1994." -> "1994")
-            if final_answer.endswith('.'):
-                final_answer = final_answer[:-1]
-                
-            # Toglie virgolette extra o asterischi di formattazione Markdown
-            final_answer = final_answer.replace("**", "").replace('"', "").replace("'", "")
-            
-            print(f"[RISPOSTA PULITA TROVATA]: {final_answer}")
-            return final_answer
+            final_answer = self._clean_answer(raw_answer)
+
+            if final_answer and final_answer.lower() not in ["error", "none", "n/a", ""]:
+                print(f"[✅ RISPOSTA AGENTE]: {final_answer}")
+                return final_answer
+            else:
+                print("[⚠️ Agente ha restituito risposta vuota/nulla, provo fallback...]")
         except Exception as e:
-            print(f"Errore durante l'elaborazione: {e}")
-            return "Error"
+            print(f"[⚠️ AGENT ERROR]: {e} — provo fallback diretto...")
+
+        # --- TENTATIVO 2: Modello diretto senza tools ---
+        fallback_answer = self._direct_answer(processed_question)
+        print(f"[🔄 FALLBACK RISPOSTA]: {fallback_answer}")
+        return fallback_answer
 
 
 # ==========================================
 # ⚙️ INTERFACCIA E RUNNER
 # ==========================================
 def run_and_submit_all(profile: gr.OAuthProfile | None):
-    space_id = os.getenv("SPACE_ID") 
+    space_id = os.getenv("SPACE_ID")
 
     if profile:
         username = f"{profile.username}"
@@ -122,7 +298,7 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
         agent = SuperAgent()
     except Exception as e:
         return f"Errore nell'inizializzazione dell'agente: {e}", None
-        
+
     agent_code = f"https://huggingface.co/spaces/{space_id}/tree/main"
 
     try:
@@ -130,31 +306,44 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
         response.raise_for_status()
         questions_data = response.json()
         if not questions_data:
-             return "La lista delle domande scaricata è vuota.", None
+            return "La lista delle domande scaricata è vuota.", None
     except Exception as e:
         return f"Errore nel download delle domande: {e}", None
 
     results_log = []
     answers_payload = []
-    
+
     print(f"Avvio elaborazione su {len(questions_data)} domande...")
     for item in questions_data:
-        task_id = item.get("task_id")
+        task_id = item.get("task_id", "")
         question_text = item.get("question")
         if not task_id or question_text is None:
             continue
         try:
-            submitted_answer = agent(question_text)
+            submitted_answer = agent(question_text, task_id=task_id)
             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})
+            results_log.append({
+                "Task ID": task_id,
+                "Question": question_text[:120],
+                "Submitted Answer": submitted_answer
+            })
         except Exception as e:
-             results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": f"AGENT ERROR: {e}"})
+            answers_payload.append({"task_id": task_id, "submitted_answer": "I don't know"})
+            results_log.append({
+                "Task ID": task_id,
+                "Question": question_text[:120],
+                "Submitted Answer": f"AGENT ERROR: {e}"
+            })
 
     if not answers_payload:
         return "L'agente non ha prodotto risposte da inviare.", pd.DataFrame(results_log)
 
-    submission_data = {"username": username.strip(), "agent_code": agent_code, "answers": answers_payload}
-    
+    submission_data = {
+        "username": username.strip(),
+        "agent_code": agent_code,
+        "answers": answers_payload
+    }
+
     try:
         response = requests.post(submit_url, json=submission_data, timeout=60)
         response.raise_for_status()
@@ -171,6 +360,7 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
         status_message = f"❌ Invio Fallito: {e}"
         return status_message, pd.DataFrame(results_log)
 
+
 # --- Build Gradio Interface ---
 with gr.Blocks() as demo:
     gr.Markdown("# 🚀 Super Agente - Final Assignment Runner")
@@ -185,6 +375,4 @@ with gr.Blocks() as demo:
     )
 
 if __name__ == "__main__":
-    space_host_startup = os.getenv("SPACE_HOST")
-    space_id_startup = os.getenv("SPACE_ID")
     demo.launch(debug=True, share=False)