Update app.py

app.py

@@ -10,8 +10,7 @@ from transformers import AutoTokenizer, AutoModelForCausalLM, AutoModelForSequen
 sent = "dnzblgn/Sentiment-Analysis-Customer-Reviews"
 sarc = "dnzblgn/Sarcasm-Detection-Customer-Reviews"
 doc = "dnzblgn/Customer-Reviews-Classification"
-embedding_model = SentenceTransformer('multi-qa-mpnet-base-dot-v1')
-
+embedding_model = SentenceTransformer('all-MiniLM-L6-v2')  # Lightweight embedding model for CPU
 
 # Your models (no token, no fast tokenizer)
 sentiment_tokenizer = AutoTokenizer.from_pretrained("dnzblgn/Sentiment-Analysis-Customer-Reviews", use_fast=False)
@@ -23,13 +22,9 @@ sarcasm_model = AutoModelForSequenceClassification.from_pretrained("dnzblgn/Sarc
 classification_tokenizer = AutoTokenizer.from_pretrained("dnzblgn/Customer-Reviews-Classification", use_fast=False)
 classification_model = AutoModelForSequenceClassification.from_pretrained("dnzblgn/Customer-Reviews-Classification")
 
-# Mistral model (requires token, must be authenticated)
-HF_TOKEN = os.getenv("rag")  # Using the secret from your Hugging Face Space
-if HF_TOKEN is None:
-    raise ValueError("Environment variable 'rag' is not set. Please check your Space secrets.")
-
-mistral_tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1", use_fast=False, use_auth_token=HF_TOKEN)
-mistral_model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1", use_auth_token=HF_TOKEN).eval()
+# Lightweight Causal Language Model (distilgpt2 instead of Mistral)
+causal_tokenizer = AutoTokenizer.from_pretrained("distilgpt2")
+causal_model = AutoModelForCausalLM.from_pretrained("distilgpt2").eval()  # Ensure evaluation mode
 
 # Paths and files
 UPLOAD_FOLDER = "uploads"
@@ -133,11 +128,11 @@ def handle_uploaded_file(file):
 
     return "File uploaded and processed successfully."
 
-def mistral_generate_response(prompt):
-    inputs = mistral_tokenizer(prompt, return_tensors="pt")  # Default is CPU
+def causal_generate_response(prompt):
+    inputs = causal_tokenizer(prompt, return_tensors="pt")  # Default CPU
     with torch.no_grad():
-        outputs = mistral_model.generate(inputs["input_ids"], max_length=500, do_sample=True, temperature=0.7)
-    response = mistral_tokenizer.decode(outputs[0], skip_special_tokens=True)
+        outputs = causal_model.generate(inputs["input_ids"], max_length=500, do_sample=True, temperature=0.7)
+    response = causal_tokenizer.decode(outputs[0], skip_special_tokens=True)
     return response
 
 def query_chatbot(query):
@@ -153,13 +148,18 @@ def query_chatbot(query):
     relevant_docs = [documents[idx] for idx in indices[0] if idx < len(documents)]
     context = "\n\n".join(relevant_docs[:top_k])
 
+    # Custom Prompt for RAG
     final_prompt = (
+        f"You are a business data analyst. Analyze the feedback data and identify the overall sentiment trends. "
+        f"Focus on determining whether positive feedback or negative feedback dominates in each category, and avoid overstating less significant trends. "
+        f"Provide clear, data-driven insights.\n\n"
         f"Context:\n{context}\n\n"
         f"Question: {query}\n\n"
-        f"Your Answer (based on the context):"
+        f"Your Answer (based on the data and context):"
     )
 
-    return mistral_generate_response(final_prompt)
+    return causal_generate_response(final_prompt)
+
 
 # Gradio interface
 with gr.Blocks() as interface:
@@ -190,4 +190,4 @@ with gr.Blocks() as interface:
 
 # Run Gradio app
 if __name__ == "__main__":
-    interface.launch()
+    interface.launch()