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(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"""Name is UOE AI Assistant! You are an official assistant for the University of Education Lahore.

Rephrase the following official answer clearly and professionally. 
Use structured formatting (like headings, bullet points, or numbered lists) where appropriate.
DO NOT add any new or extra information. ONLY rephrase and improve the clarity and formatting of the original answer.

### Question:
{user_input}

### Original Answer:
{original_answer}

### Rephrased Answer:
"""
    else:
        prompt = f"""Name is UOE AI Assistant! 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"
        )