Update app.py

app.py

@@ -10,35 +10,76 @@ DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 
 # --- Tool Implementations ---
 
-def download_task_file(task_id: str) -> str:
-    """Checks the API for attached files and downloads them to the working directory."""
+def download_and_read_task_file(task_id: str):
+    """Downloads file and immediately reads its content. Returns (filename, content_str)."""
     url = f"{DEFAULT_API_URL}/files/{task_id}"
     try:
         response = requests.get(url, timeout=15)
-        if response.status_code == 200:
-            cd = response.headers.get('content-disposition')
-            filename = f"file_{task_id[:8]}.tmp"
-            if cd:
-                match = re.search(r'filename="?([^"]+)"?', cd)
-                if match:
-                    filename = match.group(1)
-            with open(filename, 'wb') as f:
-                f.write(response.content)
-            return (
-                f"\n[SYSTEM: File '{filename}' downloaded. "
-                f"READ IT FIRST using PYTHON:\n```python\n"
-                f"try:\n"
-                f"    with open('{filename}', 'r', encoding='utf-8') as f:\n"
-                f"        print(f.read())\n"
-                f"except:\n"
-                f"    with open('{filename}', 'rb') as f:\n"
-                f"        print(repr(f.read()[:2000]))\n"
-                f"```\n]\n"
-            )
-        return ""
+        if response.status_code != 200:
+            return None, ""
+
+        cd = response.headers.get('content-disposition', '')
+        filename = f"file_{task_id[:8]}.tmp"
+        match = re.search(r'filename="?([^"]+)"?', cd)
+        if match:
+            filename = match.group(1)
+
+        with open(filename, 'wb') as f:
+            f.write(response.content)
+
+        print(f"  [File downloaded: {filename}]")
+
+        # Try to read content immediately based on file type
+        ext = filename.lower().split('.')[-1]
+
+        if ext in ['xlsx', 'xls']:
+            try:
+                df = pd.read_excel(filename, sheet_name=None)
+                content = ""
+                for sheet, data in df.items():
+                    content += f"Sheet: {sheet}\n{data.to_string()}\n\n"
+                return filename, content[:3000]
+            except Exception as e:
+                return filename, f"Excel read error: {e}"
+
+        elif ext == 'py':
+            try:
+                with open(filename, 'r', encoding='utf-8') as f:
+                    return filename, f.read()
+            except Exception as e:
+                return filename, f"Python file read error: {e}"
+
+        elif ext in ['txt', 'csv', 'json', 'md']:
+            try:
+                with open(filename, 'r', encoding='utf-8') as f:
+                    return filename, f.read()[:3000]
+            except Exception as e:
+                return filename, f"Text read error: {e}"
+
+        elif ext in ['mp3', 'wav', 'ogg', 'm4a']:
+            # Audio - try whisper if available, else note it
+            try:
+                import whisper
+                model = whisper.load_model("tiny")
+                result = model.transcribe(filename)
+                return filename, f"Audio transcript: {result['text']}"
+            except Exception:
+                return filename, f"Audio file '{filename}' downloaded but cannot be transcribed (no whisper). File size: {len(response.content)} bytes."
+
+        elif ext in ['png', 'jpg', 'jpeg', 'gif', 'webp']:
+            return filename, f"Image file '{filename}' downloaded. Size: {len(response.content)} bytes. Cannot read image content directly."
+
+        else:
+            # Try text first, fall back to binary
+            try:
+                with open(filename, 'r', encoding='utf-8') as f:
+                    return filename, f.read()[:3000]
+            except Exception:
+                return filename, f"Binary file '{filename}' downloaded. Size: {len(response.content)} bytes."
+
     except Exception as e:
-        print(f"File download error for {task_id}: {e}")
-        return ""
+        print(f"  File download error for {task_id}: {e}")
+        return None, ""
 
 
 def web_search(query: str) -> str:
@@ -50,7 +91,7 @@ def web_search(query: str) -> str:
             return "No results found."
         output = []
         for r in results:
-            output.append(f"Title: {r.get('title','')}\nURL: {r.get('href','')}\nSnippet: {r.get('body','')[:200]}")
+            output.append(f"Title: {r.get('title','')}\nURL: {r.get('href','')}\nSnippet: {r.get('body','')[:300]}")
         return "\n---\n".join(output)
     except ImportError:
         try:
@@ -61,7 +102,7 @@ def web_search(query: str) -> str:
                 return "No results found."
             output = []
             for r in results:
-                output.append(f"Title: {r.get('title','')}\nURL: {r.get('href','')}\nSnippet: {r.get('body','')[:200]}")
+                output.append(f"Title: {r.get('title','')}\nURL: {r.get('href','')}\nSnippet: {r.get('body','')[:300]}")
             return "\n---\n".join(output)
         except Exception as e:
             return f"Search error: {e}"
@@ -70,7 +111,6 @@ def web_search(query: str) -> str:
 
 
 def web_fetch(url: str) -> str:
-    """Fetches the text content of a web page."""
     try:
         headers = {"User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36"}
         response = requests.get(url, timeout=20, headers=headers)
@@ -140,7 +180,6 @@ class SmartAgent:
         print(f"SmartAgent initialized with Groq ({self.model})")
 
     def call_llm(self, prompt: str) -> str:
-        # Hard cap prompt to avoid 413
         if len(prompt) > 8000:
             prompt = prompt[:3000] + "\n\n[...context trimmed...]\n\n" + prompt[-3000:]
 
@@ -175,35 +214,43 @@ class SmartAgent:
     def __call__(self, question: str, task_id: str) -> str:
         print(f"\nQuestion: {question[:100]}...")
 
-        file_alert = download_task_file(task_id)
+        # Download and immediately read any attached file
+        filename, file_content = download_and_read_task_file(task_id)
+
+        file_context = ""
+        if filename and file_content:
+            file_context = f"\n\n[ATTACHED FILE: '{filename}']\n{file_content}\n[END OF FILE]"
+        elif filename:
+            file_context = f"\n\n[ATTACHED FILE: '{filename}' - could not read content]"
 
         system = """You are an expert AI solving benchmark questions. Think step by step.
 
 TOOLS AVAILABLE:
-- SEARCH: <query>          (web search)
-- WIKIPEDIA: <query>       (Wikipedia lookup)
-- FETCH: <url>             (read a webpage)
-- PYTHON: ```python ... ``` (run code, always use print())
-
-ANSWER FORMAT: When done, output exactly:
-ANSWER: <your answer>
-
-RULES:
-- Read any downloaded file FIRST with PYTHON
-- Use FETCH to read URLs mentioned in questions
-- Be concise - one tool per step
-- Numbers: no units unless asked, no commas
-- Short phrases, no articles (a/the)
-- Comma-separated if multiple items needed"""
-
-        # Keep a short rolling history instead of full conversation
+- SEARCH: <query>
+- WIKIPEDIA: <query>
+- FETCH: <url>
+- PYTHON: ```python ... ```  (always use print())
+
+OUTPUT FORMAT:
+THOUGHT: <reasoning>
+SEARCH: <query>
+
+Or when done:
+ANSWER: <exact answer>
+
+STRICT RULES:
+- If file content is provided above, use it directly - DO NOT re-read it
+- Answer must be exact: numbers only (no units unless asked), short phrases
+- No articles (a/the), no commas in numbers
+- Comma-separated list if multiple items needed
+- Do NOT say "Unable to determine" - always give your best guess"""
+
         history = []
-        initial_prompt = f"{system}\n\nQuestion: {question}{file_alert}"
+        initial_prompt = f"{system}\n\nQuestion: {question}{file_context}"
 
-        for iteration in range(6):
+        for iteration in range(8):
             time.sleep(15)
 
-            # Build prompt from system + question + last 3 exchanges only
             if not history:
                 prompt = initial_prompt
             else:
@@ -212,19 +259,22 @@ RULES:
                     f"Step {i+1}:\nAction: {h['action']}\nResult: {h['result'][:400]}"
                     for i, h in enumerate(recent)
                 ])
-                prompt = f"{system}\n\nQuestion: {question}{file_alert}\n\nPrevious steps:\n{exchanges}\n\nContinue:"
+                prompt = f"{system}\n\nQuestion: {question}{file_context}\n\nPrevious steps:\n{exchanges}\n\nContinue:"
 
             response = self.call_llm(prompt)
             print(f"  LLM [{iteration}]: {response[:200]}...")
 
-            # Parse tool calls
             fetch_match = re.search(r'FETCH:\s*(https?://\S+)', response)
             search_match = re.search(r'SEARCH:\s*(.+?)(?:\n|$)', response)
             wiki_match = re.search(r'WIKIPEDIA:\s*(.+?)(?:\n|$)', response)
             python_match = re.search(r'PYTHON:\s*```(?:python)?\n?(.*?)```', response, re.DOTALL)
             answer_match = re.search(r'ANSWER:\s*(.+?)(?:\n|$)', response, re.IGNORECASE)
 
-            if fetch_match:
+            if answer_match:
+                answer = answer_match.group(1).strip()
+                print(f"  Final Answer: {answer}")
+                return answer
+            elif fetch_match:
                 url = fetch_match.group(1).strip()
                 print(f"  Tool: FETCH({url[:80]})")
                 result = web_fetch(url)
@@ -244,17 +294,18 @@ RULES:
                 print(f"  Tool: PYTHON({code[:60]}...)")
                 result = run_python(code)
                 history.append({"action": f"PYTHON: {code[:100]}", "result": result})
-            elif answer_match:
-                answer = answer_match.group(1).strip()
-                print(f"  Final Answer: {answer}")
-                return answer
             else:
-                history.append({"action": "none", "result": "No valid tool found. Use SEARCH, WIKIPEDIA, FETCH, PYTHON, or ANSWER."})
+                history.append({"action": "none", "result": "No valid tool. Use SEARCH, WIKIPEDIA, FETCH, PYTHON, or ANSWER."})
 
-        # Fallback: force answer
+        # Fallback: force a best-guess answer
         recent = history[-3:]
         exchanges = "\n\n".join([f"Action: {h['action']}\nResult: {h['result'][:300]}" for h in recent])
-        fallback_prompt = f"{system}\n\nQuestion: {question}\n\nSteps taken:\n{exchanges}\n\nOut of steps. Give ONLY: ANSWER: <answer>"
+        fallback_prompt = (
+            f"{system}\n\nQuestion: {question}{file_context}\n\n"
+            f"Steps taken:\n{exchanges}\n\n"
+            f"You MUST give a final answer now. Do not say 'unable to determine'. "
+            f"Use your best judgment. Output ONLY: ANSWER: <answer>"
+        )
         last = self.call_llm(fallback_prompt)
         answer_match = re.search(r'ANSWER:\s*(.+?)(?:\n|$)', last, re.IGNORECASE)
         if answer_match:
@@ -343,7 +394,7 @@ with gr.Blocks() as demo:
     gr.Markdown("# 🤖 Smart Agent — GAIA Benchmark Runner")
     gr.Markdown(
         """
-        **Powered by Groq (Llama 3.1 8B Instant)**
+        **Powered by Groq (Llama 3.3 70B)**
         **Instructions:**
         1. Make sure `GROQ_API_KEY` is set in your Space secrets
         2. `requirements.txt` must include: `gradio`, `requests`, `pandas`, `openpyxl`, `ddgs`, `beautifulsoup4`