Update app.py

app.py

@@ -7,7 +7,7 @@ import gradio as gr
 
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 
-# ── Imports exacts depuis la doc officielle smolagents ──────────────────
+# ── Imports officiels smolagents ──────────────────────────────────────
 from smolagents import (
     CodeAgent,
     InferenceClientModel,
@@ -16,7 +16,47 @@ from smolagents import (
     tool,
 )
 
-# ── Tools custom ────────────────────────────────────────────────────────
+# ── Prompt templates COMPLETS (obligatoires pour CodeAgent) ───────────
+def get_prompt_templates():
+    return {
+        "system_prompt": """You are an expert AI assistant solving GAIA benchmark tasks.
+You have access to tools and must use them to find accurate answers.
+
+RULES:
+- Always use Thought: then Code: sequences
+- Return ONLY the exact answer - no explanation
+- For reversed text: reverse it back then answer  
+- For math/logic: write Python code to compute
+- For files: use the download tools
+- Answers are exact-match graded
+
+{{authorized_imports}}
+""",
+        "planning": """
+Facts given in the task:
+<<facts_given_in_task>>
+
+Facts needed:
+<<facts_needed>>
+
+Plan:
+<<plan>>
+
+<end_plan>
+""",
+        "managed_agent": """
+You are a managed agent. Return your result via final_answer().
+Task: {{task}}
+""",
+        "final_answer": """
+Return ONLY the final answer. No explanation. No punctuation unless required.
+- Numbers: digits only (e.g. 42)
+- Lists: comma-separated (e.g. apple, banana)
+- Names: as-is
+"""
+    }
+
+# ── Tools custom ──────────────────────────────────────────────────────
 
 @tool
 def wikipedia_search(query: str) -> str:
@@ -26,17 +66,17 @@ def wikipedia_search(query: str) -> str:
     """
     try:
         base = "https://en.wikipedia.org/w/api.php"
-        search = requests.get(base, params={
+        r = requests.get(base, params={
             "action": "query", "list": "search",
             "srsearch": query, "format": "json", "srlimit": 1,
         }, timeout=15).json()
-        title = search["query"]["search"][0]["title"]
-        extract = requests.get(base, params={
+        title = r["query"]["search"][0]["title"]
+        ex = requests.get(base, params={
             "action": "query", "prop": "extracts",
             "exintro": True, "explaintext": True,
             "titles": title, "format": "json",
         }, timeout=15).json()
-        pages = extract["query"]["pages"]
+        pages = ex["query"]["pages"]
         text = next(iter(pages.values())).get("extract", "")[:4000]
         return f"# {title}\n{text}"
     except Exception as e:
@@ -73,8 +113,7 @@ def download_file_for_task(task_id: str) -> str:
             for _ in range(3):
                 resp = requests.post(url, headers={"Authorization": f"Bearer {token}"}, data=data, timeout=120)
                 if resp.status_code == 503:
-                    time.sleep(20)
-                    continue
+                    time.sleep(20); continue
                 if resp.status_code == 200:
                     return resp.json().get("text", "")
             return "Audio transcription failed."
@@ -102,15 +141,14 @@ def download_file_for_task(task_id: str) -> str:
                 "model": "meta-llama/Llama-3.2-11B-Vision-Instruct",
                 "messages": [{"role": "user", "content": [
                     {"type": "image_url", "image_url": {"url": f"data:{mime};base64,{b64}"}},
-                    {"type": "text", "text": "Describe everything in this image in detail. If it's a chess board, name every piece and its exact square. If there is text or numbers, transcribe them exactly."},
+                    {"type": "text", "text": "Describe everything in detail. If chess: name every piece and square. Transcribe any text/numbers exactly."},
                 ]}],
                 "max_tokens": 1024,
             }
             for _ in range(3):
                 resp = requests.post(url, headers={"Authorization": f"Bearer {token}", "Content-Type": "application/json"}, json=payload, timeout=120)
                 if resp.status_code == 503:
-                    time.sleep(20)
-                    continue
+                    time.sleep(20); continue
                 if resp.status_code == 200:
                     return resp.json()["choices"][0]["message"]["content"]
             return "Image analysis failed."
@@ -126,7 +164,7 @@ def download_file_for_task(task_id: str) -> str:
 def get_youtube_transcript(video_url: str) -> str:
     """Fetch the transcript/captions from a YouTube video URL.
     Args:
-        video_url: The full YouTube URL, e.g. https://www.youtube.com/watch?v=XXXXX
+        video_url: The full YouTube URL e.g. https://www.youtube.com/watch?v=XXXXX
     """
     try:
         from youtube_transcript_api import YouTubeTranscriptApi
@@ -141,7 +179,7 @@ def get_youtube_transcript(video_url: str) -> str:
 
 @tool
 def run_python_code(code: str) -> str:
-    """Execute Python code and return stdout. Use for math, logic, string manipulation, data processing.
+    """Execute Python code and return stdout. Use for math, logic, string ops, data processing.
     Args:
         code: Valid Python code to execute.
     """
@@ -155,11 +193,10 @@ def run_python_code(code: str) -> str:
         return f"Execution error: {e}"
 
 
-# ── Agent ───────────────────────────────────────────────────────────────
+# ── Agent ─────────────────────────────────────────────────────────────
 
 class GAIAAgent:
     def __init__(self):
-        # Exactement comme dans la doc officielle smolagents
         model = InferenceClientModel(
             model_id="meta-llama/Llama-3.3-70B-Instruct",
             token=os.environ.get("HF_TOKEN", ""),
@@ -174,11 +211,13 @@ class GAIAAgent:
                 run_python_code,
             ],
             model=model,
+            add_base_tools=True,
             max_steps=10,
             verbosity_level=1,
             additional_authorized_imports=[
                 "re", "json", "math", "unicodedata",
                 "datetime", "collections", "itertools",
+                "pandas", "requests", "os", "time",
             ],
         )
         print("GAIAAgent ready ✅")
@@ -186,34 +225,25 @@ class GAIAAgent:
     def __call__(self, question: str, task_id: str = "") -> str:
         print(f"\n{'='*60}\nQ: {question[:120]}")
 
-        # Détection de fichier joint ou YouTube dans la question
-        has_yt = bool(re.search(r"youtube\.com|youtu\.be", question))
-        has_file_hint = any(w in question.lower() for w in ["attached", "file", "image", "audio", "excel", "spreadsheet", "pdf", "code"])
-
         task_hint = ""
-        if task_id and (has_file_hint or has_yt):
-            task_hint = f"\n\nNote: task_id='{task_id}' — use download_file_for_task('{task_id}') if a file is needed."
-        elif task_id:
-            task_hint = f"\n\n[task_id: '{task_id}' — use download_file_for_task if a file is mentioned]"
+        if task_id:
+            task_hint = f"\n\n[task_id='{task_id}' — call download_file_for_task('{task_id}') if a file/image/audio is needed]"
 
         prompt = (
-            "Solve this GAIA benchmark question. Important rules:\n"
-            "- Use tools to find/verify information. Do NOT guess.\n"
-            "- For YouTube URLs → call get_youtube_transcript.\n"
-            "- For attached files (pdf/image/audio/excel/code) → call download_file_for_task.\n"
-            "- For math/logic/string manipulation → call run_python_code.\n"
-            "- For factual lookups → call wikipedia_search or DuckDuckGoSearchTool.\n"
-            "- Your final answer must be SHORT and EXACT (exact string match is used for grading).\n"
-            "- For reversed text: decode it first, then answer.\n"
-            "- For counts: give only the number.\n"
-            "- For lists: comma-separated values only.\n\n"
+            "Solve this GAIA benchmark question precisely.\n"
+            "- Use tools to verify facts. Do NOT guess.\n"
+            "- YouTube URL → call get_youtube_transcript\n"
+            "- File/image/audio/excel/pdf → call download_file_for_task\n"
+            "- Math/logic/strings → call run_python_code\n"
+            "- Facts → wikipedia_search or DuckDuckGoSearchTool\n"
+            "- Reversed text → decode first, then answer\n"
+            "- Return ONLY the exact answer. No explanation.\n\n"
             f"Question: {question}{task_hint}"
         )
 
         try:
             result = self.agent.run(prompt)
             answer = str(result).strip()
-            # Nettoyer les préfixes verbeux du LLM
             for prefix in ["the answer is", "answer:", "final answer:", "result:"]:
                 if answer.lower().startswith(prefix):
                     answer = answer[len(prefix):].strip().lstrip(":").strip()
@@ -224,7 +254,7 @@ class GAIAAgent:
             return "Unable to determine answer."
 
 
-# ── Gradio UI ────────────────────────────────────────────────────────────
+# ── Gradio UI ─────────────────────────────────────────────────────────
 
 def run_and_submit_all(profile: gr.OAuthProfile | None):
     if not profile:
@@ -257,7 +287,7 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
         answer = agent(question_text, task_id=task_id)
         answers_payload.append({"task_id": task_id, "submitted_answer": answer})
         results_log.append({"Task ID": task_id, "Question": question_text[:80], "Submitted Answer": answer})
-        time.sleep(1)
+        time.sleep(2)
 
     if not answers_payload:
         return "No answers produced.", pd.DataFrame(results_log)
@@ -287,8 +317,8 @@ with gr.Blocks() as demo:
     gr.Markdown("# 🤖 GAIA Agent — smolagents + HF Inference")
     gr.Markdown("""
 **Models:** Llama-3.3-70B · Llama-3.2-11B-Vision · Whisper large-v3  
-**Tools:** DuckDuckGo · Wikipedia · VisitWebpage · YouTube transcript · Python · File reader (PDF/Excel/Audio/Image)  
-**Setup:** Ajoute `HF_TOKEN` dans les secrets de ton Space.
+**Tools:** DuckDuckGo · Wikipedia · VisitWebpage · YouTube transcript · Python · File reader  
+**Setup:** Ajoute `HF_TOKEN` dans les secrets de ton Space HF.
     """)
     gr.LoginButton()
     run_btn = gr.Button("🚀 Run Evaluation & Submit All Answers", variant="primary")