rag.py

import json
from sentence_transformers import SentenceTransformer, util
from groq import Groq
from datetime import datetime
import os
import pandas as pd
from datasets import load_dataset, Dataset
from dotenv import load_dotenv

# Load environment variables
load_dotenv()

# Initialize Groq client
groq_client = Groq(api_key=os.getenv("GROQ_API_KEY"))

# Load similarity model
similarity_model = SentenceTransformer('paraphrase-MiniLM-L6-v2')

# Config
HF_DATASET_REPO = "midrees2806/unmatched_queries"
HF_TOKEN = os.getenv("HF_TOKEN")

# Greeting list
GREETINGS = [
    "hi", "hello", "hey", "good morning", "good afternoon", "good evening",
    "assalam o alaikum", "salam", "aoa", "hi there",
    "hey there", "greetings"
]

# Load local dataset
try:
    with open('dataset.json', 'r') as f:
        dataset = json.load(f)
    if not all(isinstance(item, dict) and 'Question' in item and 'Answer' in item for item in dataset):
        raise ValueError("Invalid dataset structure")
except Exception as e:
    print(f"Error loading dataset: {e}")
    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_questions, convert_to_tensor=True)

# Save unmatched queries to Hugging Face
def manage_unmatched_queries(query: str):
    try:
        timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
        try:
            ds = load_dataset(HF_DATASET_REPO, token=HF_TOKEN)
            df = ds["train"].to_pandas()
        except:
            df = pd.DataFrame(columns=["Query", "Timestamp", "Processed"])
        if query not in df["Query"].values:
            new_entry = {"Query": query, "Timestamp": timestamp, "Processed": False}
            df = pd.concat([df, pd.DataFrame([new_entry])], ignore_index=True)
            updated_ds = Dataset.from_pandas(df)
            updated_ds.push_to_hub(HF_DATASET_REPO, token=HF_TOKEN)
    except Exception as e:
        print(f"Failed to save query: {e}")

# Query Groq LLM
def query_groq_llm(prompt, model_name="llama3-70b-8192"):
    try:
        chat_completion = groq_client.chat.completions.create(
            messages=[{
                "role": "user",
                "content": prompt
            }],
            model=model_name,
            temperature=0.7,
            max_tokens=500
        )
        return chat_completion.choices[0].message.content.strip()
    except Exception as e:
        print(f"Error querying Groq API: {e}")
        return ""

# Main logic function to be called from Gradio
def get_best_answer(user_input):
    if not user_input.strip():
        return "Please enter a valid question."

    user_input_lower = user_input.lower().strip()

    if len(user_input_lower.split()) < 3 and not any(greet in user_input_lower for greet in GREETINGS):
        return "Please ask your question properly with at least 3 words."

    if any(greet in user_input_lower for greet in GREETINGS):
        greeting_response = query_groq_llm(
            f"You are an official assistant for University of Education Lahore. "
            f"Respond to this greeting in a friendly and professional manner: {user_input}"
        )
        return greeting_response if greeting_response else "Hello! How can I assist you today?"

    if any(keyword in user_input_lower for keyword in ["fee structure", "fees structure", "semester fees", "semester fee"]):
        return (
            "💰 For complete and up-to-date fee details for this program, we recommend visiting the official University of Education fee structure page.\n"
            "You'll find comprehensive information regarding tuition, admission charges, and other applicable fees there.\n"
            "🔗 https://ue.edu.pk/allfeestructure.php"
        )

    user_embedding = similarity_model.encode(user_input_lower, 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:
        manage_unmatched_queries(user_input)

    if best_score >= 0.65:
        original_answer = dataset_answers[best_match_idx]
        prompt = f"""As an official assistant for University of Education Lahore, provide a clear response:
        Question: {user_input}
        Original Answer: {original_answer}
        Improved Answer:"""
    else:
        prompt = f"""As an official assistant for University of Education Lahore, provide a helpful response:
        Include relevant details about university policies.
        If unsure, direct to official channels.
        Question: {user_input}
        Official Answer:"""

    llm_response = query_groq_llm(prompt)

    if llm_response:
        for marker in ["Improved Answer:", "Official Answer:", "Rephrased Answer:"]:
            if marker in llm_response:
                return llm_response.split(marker)[-1].strip()
        return llm_response
    else:
        return dataset_answers[best_match_idx] if best_score >= 0.65 else (
            "For official information:\n"
            "📞 +92-42-99262231-33\n"
            "✉️ info@ue.edu.pk\n"
            "🌐 https://ue.edu.pk"
        )