Create app.py

finally that have chances of going right

app.py

@@ -77,6 +77,10 @@ class BasicAgent:
 
     def __init__(self):
         print("BasicAgent initialized.")
+        # Access the globally loaded model and tokenizer
+        self.tokenizer = hf_tokenizer
+        self.model = hf_model
+
 
     def __call__(self, question: str) -> str:
         print(f"Agent received question (first 50 chars): {question[:50]}...")
@@ -93,27 +97,65 @@ class BasicAgent:
             print(f"Received {len(search_results)} search results from web_search.") # Debugging results received
 
             if search_results:
-                # Process search results to formulate an answer
-                answer_parts = []
+                # Format search results into a context string for the LLM
+                context = ""
                 for i, result in enumerate(search_results[:3]): # Use top 3 results
-                    print(f"Processing search result {i+1}: Title='{result.get('title', 'N/A')[:50]}...', Snippet='{result.get('snippet', 'N/A')[:50]}...'") # Debugging result content
+                    context += f"Result {i+1}:\n"
+                    if result.get('title'):
+                        context += f"Title: {result['title']}\n"
                     if result.get('snippet'):
-                        answer_parts.append(f"Snippet {i+1}: {result['snippet']}")
-                    elif result.get('title'):
-                         answer_parts.append(f"Result {i+1} Title: {result['title']}")
-                    # Optional: add URL
-                    # if result.get('url'):
-                    #      answer_parts.append(f"URL {i+1}: {result['url']}")
-
-                print(f"answer_parts after processing: {answer_parts}") # Debugging answer parts
-
-                if answer_parts:
-                    formulated_answer = "Based on web search:\n" + "\n".join(answer_parts)
-                    print(f"Agent returning search-based answer: {formulated_answer[:100]}...")
-                    return formulated_answer
-                else:
-                    print("Web search returned results but no useful snippets/titles found.")
-                    return "I couldn't find a specific answer from the web search results."
+                        context += f"Snippet: {result['snippet']}\n"
+                    if result.get('url'):
+                        context += f"URL: {result['url']}\n"
+                    context += "---\n"
+
+                # Construct the prompt for the LLM
+                prompt = f"Question: {question}\n\nSearch Results:\n{context}\nBased on the search results provided, please answer the question."
+                print(f"LLM Prompt (first 200 chars): {prompt[:200]}...") # Debugging prompt
+
+                try:
+                    # Encode the prompt
+                    inputs = self.tokenizer(prompt, return_tensors="pt")
+
+                    # Generate response using the LLM
+                    # Adjust generation parameters as needed
+                    output_sequences = self.model.generate(
+                        **inputs,
+                        max_length=512,  # Maximum length of the generated text
+                        num_return_sequences=1, # Number of sequences to generate
+                        no_repeat_ngram_size=2, # Avoid repeating n-grams
+                        do_sample=True, # Enable sampling
+                        top_k=50, # Sample from top_k tokens
+                        top_p=0.95, # Sample from top_p probability mass
+                        temperature=0.7, # Control randomness
+                        attention_mask=inputs['attention_mask'] # Pass attention mask
+                    )
+
+                    # Decode the generated output
+                    generated_text = self.tokenizer.decode(output_sequences[0], skip_special_tokens=True)
+                    print(f"LLM Generated Text (first 200 chars): {generated_text[:200]}...") # Debugging generated text
+
+                    # Extract the answer from the generated text
+                    # For CausalLMs like gpt2, the prompt is included in the output,
+                    # so we need to remove it.
+                    if generated_text.startswith(prompt):
+                        llm_answer = generated_text[len(prompt):].strip()
+                    else:
+                        # Fallback if the output format is unexpected
+                        llm_answer = generated_text.strip()
+
+                    if llm_answer:
+                         print(f"Agent returning LLM-based answer: {llm_answer[:100]}...")
+                         return llm_answer
+                    else:
+                         print("LLM generated empty or whitespace answer.")
+                         return "I couldn't generate a specific answer based on the search results."
+
+
+                except Exception as e:
+                    print(f"Error during LLM generation: {e}")
+                    return f"An error occurred while generating the answer using the LLM: {e}"
+
 
             else:
                 print("Web search returned no results.")
@@ -258,9 +300,10 @@ with gr.Blocks(theme=gr.themes.Soft(), title="Basic Agent Evaluation Runner") as
         **Instructions:**
         1. Ensure your agent logic is defined in the `BasicAgent` class above.
         2. **Get a SerpAPI key and add it as an environment variable in your runtime environment (e.g., as a secret in your Hugging Face Space settings).**
-        3. Log in to Hugging Face using the button below.
-        4. Click the "Run Evaluation & Submit All Answers" button.
-        5. The application will fetch questions, run your agent, submit answers, and display the results below.
+        3. **Ensure your Hugging Face model and tokenizer are loaded (usually in a preceding cell).**
+        4. Log in to Hugging Face using the button below.
+        5. Click the "Run Evaluation & Submit All Answers" button.
+        6. The application will fetch questions, run your agent, submit answers, and display the results below.
         """
     )
     login_btn = gr.LoginButton()