refactor: improve response

app.py

@@ -17,7 +17,7 @@ DATA_DIR = "/data" if os.path.exists("/data") else "."
 DATASET_DIR = os.path.join(DATA_DIR, "rag_dataset")
 DATASET_PATH = os.path.join(DATASET_DIR, "dataset")
 TOKENIZER_MODEL = "google/flan-t5-small"
-SUMMARIZATION_MODEL= "Falconsai/text_summarization"
+SUMMARIZATION_MODEL= "tiiuae/falcon-40b"
 
 @st.cache_resource
 def load_local_model():
@@ -209,44 +209,47 @@ def generate_answer(question, context, max_length=512):
     clean_question = clean_text(question)
     
     # Format the input for T5 (it expects a specific format)
-    input_text = f"""Generate a detailed and well-structured answer about autism, using the provided research papers as references to support your explanations.
+    input_text = f"""Objective:
+Generate a clear, informative, and well-structured answer about autism, making the content easy to understand for a general audience. Use the provided research papers to support your explanations.
 
 Question: {clean_question}
 
-Research Papers:  
+Research Papers:
 {clean_context}
 
-Instructions:  
-1. Begin with a clear and concise overview of autism, explaining its key characteristics and significance.  
-2. Use the research papers to support your explanation, citing them in the format: "According to [PAPER TITLE], ...".  
-3. Integrate findings from the papers naturally into your response, ensuring the information is accurate and relevant.  
-4. Focus on providing informative, helpful, and easy-to-understand insights.  
+Instructions:
 
-Write your answer in a professional and accessible tone, ensuring it is well-organized and grounded in the provided research."""
+Start with a simple explanation – Clearly define what autism is in an easy-to-understand way, avoiding overly complex technical terms.
+Use real-life examples – Whenever possible, include practical examples to illustrate key concepts.
+Cite research in an accessible way – Instead of just referencing papers, explain their findings in a way that anyone can understand. Example: "A study from X University found that..."
+Avoid scientific jargon – If a technical term is necessary, provide a simple explanation.
+Organize the response into sections – Use lists and short paragraphs to improve readability.
+Write your answer in a friendly and accessible tone, ensuring that anyone, regardless of their background knowledge, can understand the information provided."""
     
     try:
         # T5 expects a specific format for the input
         inputs = tokenizer(input_text, 
-                         return_tensors="pt", 
-                         max_length=1024, 
-                         truncation=True,
-                         padding=True)
-        
+                    return_tensors="pt", 
+                    max_length=1024, 
+                    truncation=True,
+                    padding=True)
+
         with torch.inference_mode():
             outputs = model.generate(
                 **inputs,
                 max_length=max_length,
                 min_length=200,
-                num_beams=5,
-                length_penalty=1.5,
-                temperature=0.7,
+                num_beams=3,  # Reduzindo para mais variedade
+                length_penalty=1.2,  # Melhor equilíbrio entre concisão e detalhes
+                temperature=0.8,  # Aumentando um pouco para mais fluidez
                 repetition_penalty=1.2,
                 early_stopping=True,
-                no_repeat_ngram_size=3,
+                no_repeat_ngram_size=2,  # Mantendo variação no texto
                 do_sample=True,
-                top_k=50,
-                top_p=0.95
+                top_k=30,  # Reduzindo para respostas mais coerentes
+                top_p=0.9  # Equilibrando diversidade e precisão
             )
+
         
         response = tokenizer.decode(outputs[0], skip_special_tokens=True)
         response = clean_text(response)