Update app.py

app.py

@@ -8,13 +8,19 @@ import gradio as gr
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 
 # ──────────────────────────────────────────────
-# Tools
+# smolagents — InferenceClientModel est dans smolagents directement
 # ──────────────────────────────────────────────
+from smolagents import (
+    tool,
+    CodeAgent,
+    InferenceClientModel,
+    DuckDuckGoSearchTool,
+    VisitWebpageTool,
+)
 
-from huggingface_hub import InferenceClient
-from smolagents import tool, CodeAgent
-from smolagents import DuckDuckGoSearchTool, VisitWebpageTool
-
+# ──────────────────────────────────────────────
+# Tools
+# ──────────────────────────────────────────────
 
 @tool
 def wikipedia_search(query: str) -> str:
@@ -66,15 +72,14 @@ def download_file_for_task(task_id: str) -> str:
             except Exception as e:
                 return f"PDF error: {e}"
 
-        # Audio → transcribe
-        if any(x in ct for x in ["audio", "mpeg", "mp3", "wav", "ogg"]) or data[:3] == b"ID3":
+        # Audio
+        if any(x in ct for x in ["audio", "mpeg", "mp3", "wav"]) or data[:3] == b"ID3":
             hf_url = "https://api-inference.huggingface.co/models/openai/whisper-large-v3"
             headers = {"Authorization": f"Bearer {os.environ.get('HF_TOKEN','')}"}
             for _ in range(3):
                 resp = requests.post(hf_url, headers=headers, 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."
@@ -94,16 +99,10 @@ def download_file_for_task(task_id: str) -> str:
                     return f"Spreadsheet error: {e}"
 
         # Image → vision model
-        if any(x in ct for x in ["image", "png", "jpg", "jpeg", "gif", "webp"]):
+        if any(x in ct for x in ["image", "png", "jpg", "jpeg"]):
             import base64
             b64 = base64.b64encode(data).decode()
-            # Detect media type
-            if data[:4] == b"\x89PNG":
-                mime = "image/png"
-            elif data[:2] == b"\xff\xd8":
-                mime = "image/jpeg"
-            else:
-                mime = "image/png"
+            mime = "image/png" if data[:4] == b"\x89PNG" else "image/jpeg"
             hf_url = "https://api-inference.huggingface.co/models/meta-llama/Llama-3.2-11B-Vision-Instruct/v1/chat/completions"
             headers = {
                 "Authorization": f"Bearer {os.environ.get('HF_TOKEN','')}",
@@ -111,20 +110,16 @@ def download_file_for_task(task_id: str) -> str:
             }
             payload = {
                 "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 this image in detail. If it's a chess position, describe every piece and its square. If it contains text or numbers, transcribe them exactly."},
-                    ]
-                }],
+                "messages": [{"role": "user", "content": [
+                    {"type": "image_url", "image_url": {"url": f"data:{mime};base64,{b64}"}},
+                    {"type": "text", "text": "Describe this image in full detail. If chess: name every piece and square. If text/numbers: transcribe exactly."},
+                ]}],
                 "max_tokens": 1024,
             }
             for _ in range(3):
                 resp = requests.post(hf_url, headers=headers, 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."
@@ -150,11 +145,10 @@ def get_youtube_transcript(video_url: str) -> str:
         if not vid_match:
             vid_match = re.search(r"youtu\.be/([A-Za-z0-9_-]{11})", video_url)
         if not vid_match:
-            return "Could not extract video ID from URL."
+            return "Could not extract video ID."
         vid_id = vid_match.group(1)
         transcript = YouTubeTranscriptApi.get_transcript(vid_id, languages=["en", "en-US", "en-GB"])
-        text = " ".join(t["text"] for t in transcript)
-        return text[:5000]
+        return " ".join(t["text"] for t in transcript)[:5000]
     except Exception as e:
         return f"Transcript error: {e}"
 
@@ -184,83 +178,59 @@ def run_python_code(code: str) -> str:
 # Agent
 # ──────────────────────────────────────────────
 
-GAIA_SYSTEM_PROMPT = """You are an expert AI solving questions from the GAIA benchmark.
-You must give EXACT answers — they are checked with exact string matching.
-
-Guidelines:
-- For reversed/encoded text: decode it yourself first before answering.
-- For Wikipedia questions: search Wikipedia with specific terms.
-- For YouTube videos: use get_youtube_transcript tool.
-- For math/logic/tables: use run_python_code tool.
-- For attached files (images, PDFs, audio, spreadsheets): use download_file_for_task tool.
-- For counting (albums, items, etc): search and count carefully.
-- Answer with ONLY the final answer — no explanation, no punctuation unless asked.
-- For lists: comma-separated unless otherwise specified.
-- For numbers: digits only unless units are asked.
-"""
-
-
-class InferenceClientModel:
-    """Thin smolagents-compatible wrapper around huggingface_hub.InferenceClient."""
-    def __init__(self, model_id: str, token: str, timeout: int = 120):
-        self.model_id = model_id
-        self.client = InferenceClient(model=model_id, token=token, timeout=timeout)
-
-    def __call__(self, messages, stop_sequences=None, **kwargs):
-        # Prepend GAIA system guidance if not already present
-        if messages and messages[0].get("role") != "system":
-            messages = [{"role": "system", "content": GAIA_SYSTEM_PROMPT}] + list(messages)
-        elif messages and messages[0].get("role") == "system":
-            messages = list(messages)
-            messages[0]["content"] = GAIA_SYSTEM_PROMPT + "\n\n" + messages[0]["content"]
-        response = self.client.chat_completion(
-            messages=messages,
-            max_tokens=2048,
-            temperature=0.1,
-            stop=stop_sequences or [],
-        )
-        return response.choices[0].message.content
-
-
-def make_agent() -> CodeAgent:
-    token = os.environ.get("HF_TOKEN", "")
-    model = InferenceClientModel(
-        model_id="meta-llama/Llama-3.3-70B-Instruct",
-        token=token,
-        timeout=120,
-    )
-    tools = [
-        DuckDuckGoSearchTool(),
-        VisitWebpageTool(),
-        wikipedia_search,
-        download_file_for_task,
-        get_youtube_transcript,
-        run_python_code,
-    ]
-    agent = CodeAgent(
-        tools=tools,
-        model=model,
-        max_steps=10,
-        verbosity_level=1,
-        additional_authorized_imports=["re", "json", "math", "datetime", "collections", "itertools"],
-    )
-    return agent
-
-
 class GAIAAgent:
     def __init__(self):
-        self.agent = make_agent()
-        print("GAIAAgent (smolagents) ready.")
+        model = InferenceClientModel(
+            model_id="meta-llama/Llama-3.3-70B-Instruct",
+            token=os.environ.get("HF_TOKEN", ""),
+            timeout=120,
+        )
+        self.agent = CodeAgent(
+            tools=[
+                DuckDuckGoSearchTool(),
+                VisitWebpageTool(),
+                wikipedia_search,
+                download_file_for_task,
+                get_youtube_transcript,
+                run_python_code,
+            ],
+            model=model,
+            max_steps=10,
+            verbosity_level=1,
+            additional_authorized_imports=[
+                "re", "json", "math", "datetime",
+                "collections", "itertools", "unicodedata",
+            ],
+        )
+        print("GAIAAgent ready ✅")
 
     def __call__(self, question: str, task_id: str = "") -> str:
-        print(f"\n{'='*55}\nQ: {question[:100]}")
+        print(f"\n{'='*55}\nQ: {question[:120]}")
+        full_question = question
+        if task_id:
+            full_question = (
+                f"{question}\n\n"
+                f"[If this question needs a file, call download_file_for_task with task_id='{task_id}']"
+            )
+        prompt = (
+            "You are solving a GAIA benchmark question. Rules:\n"
+            "- Think step by step and use tools when needed.\n"
+            "- For YouTube links: call get_youtube_transcript.\n"
+            "- For attached files (image/pdf/audio/excel): call download_file_for_task.\n"
+            "- For math/logic: call run_python_code.\n"
+            "- For facts: call wikipedia_search or DuckDuckGoSearchTool.\n"
+            "- Give ONLY the final answer, nothing else. Exact match required.\n"
+            "- For lists: comma-separated. For numbers: digits only.\n\n"
+            f"Question: {full_question}"
+        )
         try:
-            result = self.agent.run(question)
+            result = self.agent.run(prompt)
             answer = str(result).strip()
-            for prefix in ["The answer is", "Answer:", "ANSWER:", "Final answer:"]:
+            # Strip common LLM prefixes
+            for prefix in ["The answer is", "Answer:", "ANSWER:", "Final answer:", "Result:"]:
                 if answer.lower().startswith(prefix.lower()):
-                    answer = answer[len(prefix):].strip().strip(":").strip()
-            print(f"→ Answer: {answer}")
+                    answer = answer[len(prefix):].strip().lstrip(":").strip()
+            print(f"→ {answer}")
             return answer
         except Exception as e:
             print(f"Agent error: {e}")
@@ -300,7 +270,7 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
         question_text = item.get("question", "")
         if not task_id or not question_text:
             continue
-        answer = agent(question_text)
+        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,
@@ -337,7 +307,7 @@ with gr.Blocks() as demo:
     gr.Markdown("# 🤖 GAIA Agent — smolagents + HF Inference")
     gr.Markdown("""
 **Models:** Llama-3.3-70B (reasoning) · Llama-3.2-11B-Vision (images) · Whisper large-v3 (audio)  
-**Tools:** DuckDuckGo search · Wikipedia · Visit webpage · YouTube transcript · Python execution · File reader (PDF/Excel/Audio/Image)  
+**Tools:** DuckDuckGo · Wikipedia · Visit webpage · YouTube transcript · Python execution · File reader  
 **Requires:** `HF_TOKEN` secret in your Space settings.
     """)
     gr.LoginButton()