Update app.py

made changes in the paper qs answering algorithm to make it robust and accurate

app.py

@@ -79,7 +79,7 @@ def generate_embeddings_sci_bert(texts, batch_size=32):
         return np.concatenate(all_embeddings, axis=0)
     except Exception as e:
         logger.error(f"Embedding generation failed: {e}")
-        return np.zeros((len(texts), 768))  # Fallback to zero embeddings
+        return np.zeros((len(texts), 768))
 
 # Precompute embeddings and FAISS index
 try:
@@ -114,7 +114,7 @@ def get_relevant_papers(query):
         logger.error(f"Search failed: {e}")
         return [], "Search failed. Please try again."
 
-# GPT-2 QA function
+# GPT-2 QA function with direct prompting
 def answer_question(paper, question, history):
     if not paper:
         return [(question, "Please select a paper first!")], history
@@ -128,26 +128,42 @@ def answer_question(paper, question, history):
         title = paper.split(" - Abstract: ")[0].split(". ", 1)[1]
         abstract = paper.split(" - Abstract: ")[1].rstrip("...")
         
-        # Build context with history
-        context = f"Title: {title}\nAbstract: {abstract}\n\nPrevious conversation:\n"
-        for user_q, bot_a in history:
-            context += f"User: {user_q}\nAssistant: {bot_a}\n"
-        context += f"User: {question}\nAssistant: "
+        # Build a simple prompt
+        prompt = (
+            f"You are an expert assistant. Based on the following paper details:\n"
+            f"Title: {title}\n"
+            f"Abstract: {abstract}\n\n"
+            f"Answer this question: {question}"
+        )
         
-        # Generate response
-        inputs = gpt2_tokenizer(context, return_tensors="pt", truncation=True, max_length=512)
+        # Include recent history if available
+        if history:
+            prompt += "\n\nPrevious conversation:\n"
+            for user_q, bot_a in history[-2:]:  # Last 2 turns for context
+                prompt += f"User: {user_q}\nAssistant: {bot_a}\n"
+        
+        logger.info(f"Prompt sent to GPT-2: {prompt[:200]}...")
+        
+        # Generate response directly
+        inputs = gpt2_tokenizer(prompt, return_tensors="pt", truncation=True, max_length=400)
         inputs = {key: val.to(device) for key, val in inputs.items()}
         with torch.no_grad():
             outputs = gpt2_model.generate(
                 inputs["input_ids"],
-                max_new_tokens=100,
+                max_new_tokens=150,  # Longer responses for clarity
                 do_sample=True,
-                temperature=0.7,
-                top_k=50,
+                temperature=0.8,
+                top_p=0.9,
                 pad_token_id=gpt2_tokenizer.eos_token_id
             )
+        
+        # Decode full output and extract response
         response = gpt2_tokenizer.decode(outputs[0], skip_special_tokens=True)
-        response = response[len(context):].strip()
+        response = response[len(prompt):].strip()  # Remove prompt from output
+        
+        # Fallback for bad responses
+        if not response or len(response) < 10:
+            response = "I couldn’t generate a clear answer. Could you rephrase your question?"
         
         history.append((question, response))
         return history, history
@@ -218,7 +234,7 @@ with gr.Blocks(
             ).then(
                 fn=lambda: "",
                 inputs=None,
-                outputs=question_input  # Clear input
+                outputs=question_input
             )
 
 # Launch the app