Update app.py

app.py

@@ -11,42 +11,82 @@ from langchain_community.llms import HuggingFaceHub
 #     load_dotenv()
 
 # --- Constants ---
-DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
+
+# TO:
+import google.generativeai as genai # For Gemini
+
+# ... (rest of your existing imports and constants)
+
+# --- Constants ---
+DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space" # This remains the same
 
 # --- Basic Agent Definition --
 # ----- THIS IS WERE YOU CAN BUILD WHAT YOU WANT ------
 class BasicAgent:
-    def __init__(self, hf_api_token: str | None = None):
-        print("BasicAgent initializing...")
-        token_to_use = hf_api_token or os.getenv("HUGGINGFACEHUB_API_TOKEN") or os.getenv("HF_TOKEN")
+    def __init__(self, google_api_key: str | None = None): # Changed parameter name for clarity
+        print("BasicAgent initializing with Google Gemini...")
+        
+        # Determine the Google API token
+        token_to_use = google_api_key
+        if not token_to_use:
+            token_to_use = os.getenv("GOOGLE_API_KEY") # Standard environment variable for Google API keys
 
         if not token_to_use:
             raise ValueError(
-                "Hugging Face API token not found. Please set HUGGINGFACEHUB_API_TOKEN or HF_TOKEN "
-                "as a secret in your Hugging Face Space. This token is required for the LLM."
+                "Google API key not found. Please set GOOGLE_API_KEY "
+                "as a secret in your Hugging Face Space. This token is required for Gemini."
             )
         
-        self.llm_repo_id = "mistralai/Mistral-7B-Instruct-v0.1" # Or your preferred model
         try:
-            self.llm = HuggingFaceHub(
-                repo_id=self.llm_repo_id,
-                # Increased max_new_tokens as the ReAct prompt is long and might generate a longer thought process
-                # Temperature 0.0 for more deterministic ReAct output, 0.1 is also fine.
-                model_kwargs={"temperature": 0.1, "max_new_tokens": 512}, 
-                huggingfacehub_api_token=token_to_use
+            # Configure the Gemini client
+            genai.configure(api_key=token_to_use)
+            
+            # Specify the Gemini model you want to use.
+            # As of my last update, "gemini-1.5-pro-latest" is a good choice for high capability.
+            # "gemini-pro" is also a valid and generally available option.
+            # "gemini-2.5-pro" is not a known generally available model name as of now.
+            # Please use a valid model name from: https://ai.google.dev/models/gemini
+            self.model_name = "gemini-1.5-pro-latest" # Or "gemini-pro"
+            
+            # Create the GenerativeModel instance
+            self.llm = genai.GenerativeModel(self.model_name)
+            
+            # Define generation configuration (optional, but good for consistency)
+            self.generation_config = genai.types.GenerationConfig(
+                temperature=0.1, # For more deterministic output, good for agents
+                # max_output_tokens=512 # Equivalent to max_new_tokens
+                # You can set other parameters here like top_p, top_k
             )
-            print(f"BasicAgent initialized with LLM: {self.llm_repo_id}")
+            # Define safety settings (important for Gemini to avoid overly restrictive blocking)
+            # Adjust these based on your needs and if you encounter content blocking.
+            self.safety_settings = [
+                {
+                    "category": "HARM_CATEGORY_HARASSMENT",
+                    "threshold": "BLOCK_MEDIUM_AND_ABOVE" # Or BLOCK_ONLY_HIGH, BLOCK_NONE
+                },
+                {
+                    "category": "HARM_CATEGORY_HATE_SPEECH",
+                    "threshold": "BLOCK_MEDIUM_AND_ABOVE"
+                },
+                {
+                    "category": "HARM_CATEGORY_SEXUALLY_EXPLICIT",
+                    "threshold": "BLOCK_MEDIUM_AND_ABOVE"
+                },
+                {
+                    "category": "HARM_CATEGORY_DANGEROUS_CONTENT",
+                    "threshold": "BLOCK_MEDIUM_AND_ABOVE"
+                }
+            ]
+
+            print(f"BasicAgent initialized with Google Gemini model: {self.model_name}")
         except Exception as e:
-            print(f"Error initializing HuggingFaceHub: {e}")
-            raise ValueError(f"Failed to initialize LLM: {e}. Check token and model repo_id.")
+            print(f"Error initializing Google Gemini client: {e}")
+            raise ValueError(f"Failed to initialize Gemini: {e}. Check API key and model name.")
 
     # Modified signature to accept task_id (though not used in this simple version yet)
     def __call__(self, question: str, task_id: str | None = None) -> str:
-        print(f"Agent received question (Task ID: {task_id}, first 80 chars): {question[:80]}...")
+        print(f"Agent (Gemini) received question (Task ID: {task_id}, first 80 chars): {question[:80]}...")
 
-        # Prompt engineering is crucial.
-        # The `question` variable (method argument) is now correctly inserted here.
-        # This is a single-shot prompt. A true ReAct agent would have a loop.
         current_prompt = f"""You are a diligent and highly intelligent AI assistant. Your goal is to answer the given `Question` accurately and concisely.
 If the question requires multiple steps or information from tools, think step-by-step.
 
@@ -69,44 +109,181 @@ Example: If asked "What is 2+2?", your output should be "4".
 Now, please answer the following question:
 Question: {question}
 
-Answer:""" # Modified to guide the LLM towards a direct answer for this simplified agent
+Answer:"""
 
         try:
-            print(f"Sending to LLM (first 200 chars of prompt): {current_prompt[:200]}...")
-            response_text = self.llm.invoke(current_prompt)
-            answer = response_text.strip()
+            print(f"Sending to Gemini (first 200 chars of prompt): {current_prompt[:200]}...")
             
-            # Further cleaning if the model still adds prefixes or explanations
-            # This is important because we are not doing a full ReAct loop to extract "Final Answer:"
+            # FROM: response_text = self.llm.invoke(current_prompt)
+            # TO: Use Gemini's generate_content method
+            response = self.llm.generate_content(
+                current_prompt,
+                generation_config=self.generation_config,
+                safety_settings=self.safety_settings 
+            )
+
+            # Extract the text from the Gemini response object
+            # Need to handle potential errors or blocked content
+            if response.candidates:
+                if response.candidates[0].content.parts:
+                    response_text = response.candidates[0].content.parts[0].text
+                else: # Content might be empty if blocked or other issues
+                    response_text = "" 
+                    print("Warning: Gemini response has no content parts.")
+                    if response.prompt_feedback and response.prompt_feedback.block_reason:
+                        print(f"Prompt blocked by Gemini. Reason: {response.prompt_feedback.block_reason_message or response.prompt_feedback.block_reason}")
+                        return f"AGENT_ERROR: Prompt blocked by Gemini ({response.prompt_feedback.block_reason})."
+            else: # No candidates means something went wrong, possibly blocking
+                response_text = ""
+                print("Warning: Gemini response has no candidates.")
+                if response.prompt_feedback and response.prompt_feedback.block_reason:
+                    print(f"Prompt blocked by Gemini. Reason: {response.prompt_feedback.block_reason_message or response.prompt_feedback.block_reason}")
+                    return f"AGENT_ERROR: Prompt blocked by Gemini ({response.prompt_feedback.block_reason})."
+                return "AGENT_ERROR: Gemini returned no candidates in response."
+
+
+            answer = response_text.strip()
             
-            # Try to find "Answer:" if the LLM adds it despite instructions
+            # (Your existing cleaning logic can largely remain the same)
             if "Answer:" in answer:
-                # Take text after the last occurrence of "Answer:"
                 answer = answer.split("Answer:")[-1].strip()
 
-            # Remove common conversational prefixes that might slip through
             common_prefixes_to_remove = [
                 "The answer is", "My answer is", "Based on the information", "The final answer is",
                 "Here is the answer", "I found that", "It seems that"
-            ] # Case-insensitive removal
+            ]
             for prefix in common_prefixes_to_remove:
                 if answer.lower().startswith(prefix.lower()):
                     answer = answer[len(prefix):].strip()
-                    # If the first character is now a colon or period, remove it
                     if answer.startswith(":") or answer.startswith("."):
                         answer = answer[1:].strip()
-                    break # Only remove one such prefix
-
-            # If the LLM generated a ReAct-style "Final Answer:", extract from it.
+                    break 
             if "Final Answer:" in answer:
                 answer = answer.split("Final Answer:")[-1].strip()
             
-            print(f"Agent LLM raw response (first 80 chars): {response_text[:80]}...")
-            print(f"Agent cleaned answer (first 80 chars): {answer[:80]}...")
+            print(f"Agent (Gemini) LLM raw response (first 80 chars): {response_text[:80]}...")
+            print(f"Agent (Gemini) cleaned answer (first 80 chars): {answer[:80]}...")
             
             if not answer:
-                print("Warning: Agent produced an empty answer after cleaning.")
-                return "Unable to generate a valid answer."
+                print("Warning: Agent (Gemini) produced an empty answer after cleaning.")
+                return "Unable to generate a valid answer from Gemini."
+
+            return answer
+        except Exception as e:
+            # Check if it's a specific Google API error for more details
+            if hasattr(e, 'message'): # google.api_core.exceptions often have a message attribute
+                 error_message = e.message
+            else:
+                 error_message = str(e)
+            print(f"Error during Gemini LLM call for question '{question[:50]}...': {error_message}")
+            return f"AGENT_ERROR: Gemini LLM call failed. ({type(e).__name__}: {error_message})"
+
+# ... (The rest of your `run_and_submit_all` function and Gradio UI code remains unchanged) ...
+
+# At the end of the file, in the __main__ block, update the warning message:
+# if __name__ == "__main__":
+    # ... (existing startup prints) ...
+    
+    # TO: Check for GOOGLE_API_KEY
+    if not os.getenv("GOOGLE_API_KEY"):
+        print("⚠️  WARNING: GOOGLE_API_KEY environment variable not found.")
+        print("   The Gemini agent will likely fail to initialize. Please set this token in your Space secrets.")
+
+    # ... (rest of existing startup prints and demo.launch()) ...
+# --- Basic Agent Definition --
+# ----- THIS IS WERE YOU CAN BUILD WHAT YOU WANT ------
+# class BasicAgent:
+#     def __init__(self, hf_api_token: str | None = None):
+#         print("BasicAgent initializing...")
+#         token_to_use = hf_api_token or os.getenv("HUGGINGFACEHUB_API_TOKEN") or os.getenv("HF_TOKEN")
+
+#         if not token_to_use:
+#             raise ValueError(
+#                 "Hugging Face API token not found. Please set HUGGINGFACEHUB_API_TOKEN or HF_TOKEN "
+#                 "as a secret in your Hugging Face Space. This token is required for the LLM."
+#             )
+        
+#         self.llm_repo_id = "mistralai/Mistral-7B-Instruct-v0.1" # Or your preferred model
+#         try:
+#             self.llm = HuggingFaceHub(
+#                 repo_id=self.llm_repo_id,
+#                 # Increased max_new_tokens as the ReAct prompt is long and might generate a longer thought process
+#                 # Temperature 0.0 for more deterministic ReAct output, 0.1 is also fine.
+#                 model_kwargs={"temperature": 0.1, "max_new_tokens": 512}, 
+#                 huggingfacehub_api_token=token_to_use
+#             )
+#             print(f"BasicAgent initialized with LLM: {self.llm_repo_id}")
+#         except Exception as e:
+#             print(f"Error initializing HuggingFaceHub: {e}")
+#             raise ValueError(f"Failed to initialize LLM: {e}. Check token and model repo_id.")
+
+#     # Modified signature to accept task_id (though not used in this simple version yet)
+#     def __call__(self, question: str, task_id: str | None = None) -> str:
+#         print(f"Agent received question (Task ID: {task_id}, first 80 chars): {question[:80]}...")
+
+#         # Prompt engineering is crucial.
+#         # The `question` variable (method argument) is now correctly inserted here.
+#         # This is a single-shot prompt. A true ReAct agent would have a loop.
+#         current_prompt = f"""You are a diligent and highly intelligent AI assistant. Your goal is to answer the given `Question` accurately and concisely.
+# If the question requires multiple steps or information from tools, think step-by-step.
+
+# **Available Tools (Conceptual - for your reasoning process, actual tool calls are not implemented in this version):**
+
+# 1.  **`GAIAFileLookup(filename: str) -> str`**: Retrieves file content.
+# 2.  **`Calculator(expression: str) -> str`**: Performs calculations.
+# 3.  **`LLM_Query(sub_question: str) -> str`**: For general knowledge.
+
+# **Output Format Expectation:**
+# While you might reason using a "Thought:", "Action:", "Observation:" cycle internally, for this specific task, your final output should be ONLY the direct answer to the question.
+# Example: If asked "What is 2+2?", your output should be "4".
+
+# **Key Guidelines for GAIA Submission:**
+# 1.  **Conciseness:** The final answer must be precise and directly address the question.
+# 2.  **No "FINAL ANSWER" Prefix in Submission:** Do NOT include "FINAL ANSWER:" or "The answer is:" in your actual response. Just the answer value.
+
+# ---
+
+# Now, please answer the following question:
+# Question: {question}
+
+# Answer:""" # Modified to guide the LLM towards a direct answer for this simplified agent
+
+#         try:
+#             print(f"Sending to LLM (first 200 chars of prompt): {current_prompt[:200]}...")
+#             response_text = self.llm.invoke(current_prompt)
+#             answer = response_text.strip()
+            
+#             # Further cleaning if the model still adds prefixes or explanations
+#             # This is important because we are not doing a full ReAct loop to extract "Final Answer:"
+            
+#             # Try to find "Answer:" if the LLM adds it despite instructions
+#             if "Answer:" in answer:
+#                 # Take text after the last occurrence of "Answer:"
+#                 answer = answer.split("Answer:")[-1].strip()
+
+#             # Remove common conversational prefixes that might slip through
+#             common_prefixes_to_remove = [
+#                 "The answer is", "My answer is", "Based on the information", "The final answer is",
+#                 "Here is the answer", "I found that", "It seems that"
+#             ] # Case-insensitive removal
+#             for prefix in common_prefixes_to_remove:
+#                 if answer.lower().startswith(prefix.lower()):
+#                     answer = answer[len(prefix):].strip()
+#                     # If the first character is now a colon or period, remove it
+#                     if answer.startswith(":") or answer.startswith("."):
+#                         answer = answer[1:].strip()
+#                     break # Only remove one such prefix
+
+#             # If the LLM generated a ReAct-style "Final Answer:", extract from it.
+#             if "Final Answer:" in answer:
+#                 answer = answer.split("Final Answer:")[-1].strip()
+            
+#             print(f"Agent LLM raw response (first 80 chars): {response_text[:80]}...")
+#             print(f"Agent cleaned answer (first 80 chars): {answer[:80]}...")
+            
+#             if not answer:
+#                 print("Warning: Agent produced an empty answer after cleaning.")
+#                 return "Unable to generate a valid answer."
 
             return answer
         except Exception as e: