rag.py

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

# Load environment variables
load_dotenv()

# Initialize clients
groq_client = Groq(api_key=os.getenv("GROQ_API_KEY"))
similarity_model = SentenceTransformer('paraphrase-MiniLM-L6-v2')

# Configuration
HF_DATASET_REPO = "midrees2806/unmatched_queries"  # Your dataset repo
HF_TOKEN = os.getenv("HF_TOKEN")  # From Space secrets

# --- Dataset Loading ---
try:
    with open('dataset.json', 'r') as f:
        dataset = json.load(f)
    if not all(isinstance(item, dict) and 'input' in item and 'response' 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("input", "").lower().strip() for item in dataset]
dataset_answers = [item.get("response", "") for item in dataset]
dataset_embeddings = similarity_model.encode(dataset_questions, convert_to_tensor=True)

# --- Unmatched Queries Handler ---
def manage_unmatched_queries(query: str):
    """Save unmatched queries to HF Dataset with error handling"""
    try:
        timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
        
        # Load existing dataset or create new
        try:
            ds = load_dataset(HF_DATASET_REPO, token=HF_TOKEN)
            df = ds["train"].to_pandas()
        except:
            df = pd.DataFrame(columns=["Query", "Timestamp", "Processed"])
        
        # Append new query (avoid duplicates)
        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)
            
            # Push to Hub
            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}")

# --- Enhanced LLM Query ---
def query_llm(prompt: str, model: str = "llama3-70b-8192") -> str:
    try:
        response = groq_client.chat.completions.create(
            messages=[{"role": "user", "content": prompt}],
            model=model,
            temperature=0.7,
            max_tokens=1024,
            top_p=0.9
        )
        return response.choices[0].message.content.strip()
    except Exception as e:
        print(f"LLM Error: {e}")
        return None

# --- Main Chat Function ---
def get_best_answer(user_input: str) -> str:
    user_input = user_input.strip()
    lower_input = user_input.lower()

    # 1. Handle special cases
    if any(kw in lower_input for kw in ["fee", "fees", "tuition"]):
        return ("💰 Fee information:\n"
                "Please visit: https://ue.edu.pk/allfeestructure.php\n"
                "For personalized help, contact [email protected]")

    # 2. Semantic similarity search
    query_embedding = similarity_model.encode(lower_input, convert_to_tensor=True)
    scores = util.pytorch_cos_sim(query_embedding, dataset_embeddings)[0]
    best_idx = scores.argmax().item()
    best_score = scores[best_idx].item()

    # 3. Save unmatched queries (threshold = 0.65)
    if best_score < 0.65:
        manage_unmatched_queries(user_input)

    # 4. Generate response
    if best_score >= 0.65:
        context = dataset_answers[best_idx]
        prompt = f"""University Assistant Task:
        Question: {user_input}
        Context: {context}
        Generate a helpful, accurate response using the context. If unsure, say "Please contact [email protected]" """
    else:
        prompt = f"""As an official University of Education assistant, answer:
        Question: {user_input}
        Guidelines:
        - Be polite and professional
        - Direct to official channels if uncertain
        - Keep responses under 3 sentences"""

    response = query_llm(prompt)
    return response or """For official assistance:
    📞 +92-42-99262231-33
    ✉️ [email protected]"""