Update utils.py

utils.py

@@ -4,86 +4,73 @@ import requests
 from sentence_transformers import SentenceTransformer, util
 from dotenv import load_dotenv
 
-
-# ✅ Load .env file
+# Load environment variables
 load_dotenv()
 
-# ✅ Hugging Face API Token check
+# API configurations
 HF_API_TOKEN = os.getenv("HF_API_TOKEN")
-
 if not HF_API_TOKEN:
-    raise ValueError("Error: Hugging Face API Token is missing! Please check your .env file.")
-
-# ✅ Hugging Face GPT Model API Endpoint
-GPT_MODEL_API = "https://api-inference.huggingface.co/models/mistralai/Mistral-7B-Instruct-v0.2"
+    raise ValueError("Hugging Face API Token is missing from .env file")
 
-# ✅ Headers for API request
+# DeepSeek API endpoint
+DEEPSEEK_API = "https://api-inference.huggingface.co/models/deepseek-ai/deepseek-llm-7b"
 headers = {"Authorization": f"Bearer {HF_API_TOKEN}"}
 
-# ✅ Load sentence transformer model for intent matching
+# Load models
 similarity_model = SentenceTransformer('paraphrase-MiniLM-L6-v2')
 
-# ✅ Load dataset
-with open("ds.json") as f:
+# Load dataset
+with open("dataset.json") as f:
     dataset = json.load(f)
 
-# ✅ Precompute embeddings for dataset questions
-dataset_inputs = [item.get("Question", "").lower().strip() for item in dataset]
+# Precompute embeddings
+dataset_questions = [item.get("Question", "").lower().strip() for item in dataset]
 dataset_answers = [item.get("Answer", "") for item in dataset]
-dataset_embeddings = similarity_model.encode(dataset_inputs, convert_to_tensor=True)
-
-# ✅ Function to detect greetings
-def is_greeting(text):
-    greetings = ["hello", "hi", "hey", "salam", "assalam o alaikum", "assalamu alaikum", "assalamualaikum"]
-    return text.lower().strip() in greetings
-
-# ✅ Function to get the best matching answer
-def get_best_answer(user_input):
-    if is_greeting(user_input):
-        return "Hello! 😊 How can I assist you today with university-related information?"
-
-    user_input_embedding = similarity_model.encode(user_input.lower().strip(), convert_to_tensor=True)
-    similarities = util.pytorch_cos_sim(user_input_embedding, dataset_embeddings)[0]
-
-    # ✅ Find the best match and its similarity score
-    best_match_index = similarities.argmax().item()
-    best_score = similarities[best_match_index].item()
-
-    # ✅ Set a similarity threshold (tune as needed)
-    THRESHOLD = 0.65
-
-    if best_score < THRESHOLD:
-        return (
-             "I'm sorry, I couldn't find an exact answer to your question. "
-             "You may kindly try rephrasing your question gently for better results. "
-             "Also, feel free to visit the UOE official website for information: https://ue.edu.pk/"
-        )
-
-    best_answer = dataset_answers[best_match_index]
-    return rephrase_answer(best_answer)
-
-# ✅ Function to rephrase answer using GPT
-def rephrase_answer(answer):
-    prompt = (
-        f"Rephrase the following university-related answer while keeping the meaning unchanged:\n\n"
-        f"Original Answer: {answer}\n\n"
-        f"Rephrased Answer:"
-    )
-
-    payload = {"inputs": prompt}  # ✅ Structured prompt to prevent extra details
-
-    response = requests.post(GPT_MODEL_API, headers=headers, json=payload)
-
-    if response.status_code == 200:
+dataset_embeddings = similarity_model.encode(dataset_questions, convert_to_tensor=True)
+
+def query_deepseek(prompt):
+    payload = {
+        "inputs": prompt,
+        "parameters": {
+            "max_new_tokens": 500,
+            "temperature": 0.7,
+            "do_sample": True
+        }
+    }
+    try:
+        response = requests.post(DEEPSEEK_API, headers=headers, json=payload)
+        response.raise_for_status()
         result = response.json()
-        if isinstance(result, list) and result:
-            generated_text = result[0].get("generated_text", answer).strip()
-            
-            # ✅ Ensure only the rephrased answer is returned
-            if "Rephrased Answer:" in generated_text:
-                return generated_text.split("Rephrased Answer:")[-1].strip()
-            return generated_text
-
-    return answer  # ✅ Return original answer if API fails
-
+        return result[0].get("generated_text", "").strip() if isinstance(result, list) and result else ""
+    except Exception as e:
+        print(f"DeepSeek API error: {e}")
+        return ""
 
+def get_best_answer(user_input):
+    # Find best match from dataset
+    user_embedding = similarity_model.encode(user_input.lower().strip(), convert_to_tensor=True)
+    similarities = util.pytorch_cos_sim(user_embedding, dataset_embeddings)[0]
+    best_match_idx = similarities.argmax().item()
+    best_score = similarities[best_match_idx].item()
+
+    if best_score >= 0.65:  # Good match found
+        original_answer = dataset_answers[best_match_idx]
+        prompt = f"""Rephrase this university answer to be more helpful while keeping key information:
+        Question: {user_input}
+        Original Answer: {original_answer}
+        Improved Answer:"""
+    else:  # No good match
+        prompt = f"""As a university assistant, provide a helpful response to:
+        Question: {user_input}
+        Answer:"""
+
+    deepseek_response = query_deepseek(prompt)
+    
+    if deepseek_response:
+        for marker in ["Improved Answer:", "Answer:"]:
+            if marker in deepseek_response:
+                return deepseek_response.split(marker)[-1].strip()
+        return deepseek_response
+    
+    return dataset_answers[best_match_idx] if best_score >= 0.65 else """I couldn't find specific information. 
+    Please visit the UOE website: https://ue.edu.pk/"""