app.py

import streamlit as st
from openai import OpenAI
import os
import pandas as pd
import numpy as np
from sentence_transformers import SentenceTransformer
from sklearn.metrics.pairwise import cosine_similarity
import torch

# Set up OpenAI client
client = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))

# Check if GPU is available
device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Using device: {device}")

# Load metadata and embeddings (ensure these files are in your working directory or update paths)
metadata_path = 'question_metadata.csv'  # Update this path if needed
embeddings_path = 'question_dataset_embeddings.npy'  # Update this path if needed

metadata = pd.read_csv(metadata_path)
embeddings = np.load(embeddings_path)

# Load the SentenceTransformer model
model = SentenceTransformer("all-MiniLM-L6-v2").to(device)

# Load prompts from files
with open("question_generation_prompt.txt", "r") as file:
    question_generation_prompt = file.read()

with open("technical_interviewer_prompt.txt", "r") as file:
    technical_interviewer_prompt = file.read()

st.title("Real-World Programming Question Mock Interview")

# Initialize chat history in session state
if "messages" not in st.session_state:
    st.session_state.messages = []

if "follow_up_mode" not in st.session_state:
    st.session_state.follow_up_mode = False  # Tracks whether we're in follow-up mode

# Function to find the top 1 most similar question based on user input
def find_top_question(query):
    # Generate embedding for the query
    query_embedding = model.encode(query, convert_to_tensor=True, device=device).cpu().numpy()
    
    # Reshape query_embedding to ensure it is a 2D array
    query_embedding = query_embedding.reshape(1, -1)  # Reshape to (1, n_features)

    # Compute cosine similarity between query embedding and dataset embeddings
    similarities = cosine_similarity(query_embedding, embeddings).flatten()  # Flatten to get a 1D array of similarities

    # Get the index of the most similar result (top 1)
    top_index = similarities.argsort()[-1]  # Index of highest similarity

    # Retrieve metadata for the top result
    top_result = metadata.iloc[top_index].copy()
    top_result['similarity_score'] = similarities[top_index]

    return top_result

# Function to generate response using OpenAI API with debugging logs
def generate_response(messages):
    st.write("### Debugging Log: Data Sent to GPT")
    st.write(messages)  # Log the messages being sent to GPT for debugging
    
    response = client.chat.completions.create(
        model="o1-mini",
        messages=messages,
    )
    return response.choices[0].message.content

# User input form for generating a new question
with st.form(key="input_form"):
    company = st.text_input("Company", value="Google")  # Default value: Google
    difficulty = st.selectbox("Difficulty", ["Easy", "Medium", "Hard"], index=1)  # Default: Medium
    topic = st.text_input("Topic (e.g., Backtracking)", value="Backtracking")  # Default: Backtracking
    
    generate_button = st.form_submit_button(label="Generate")

if generate_button:
    # Clear session state and start fresh with follow-up mode disabled
    st.session_state.messages = []
    st.session_state.follow_up_mode = False
    
    # Create a query from user inputs and find the most relevant question
    query = f"{company} {difficulty} {topic}"
    top_question = find_top_question(query)
    
    # Prepare a detailed prompt for GPT using the top question's details
    detailed_prompt = (
        f"Transform this LeetCode question into a real-world interview scenario:\n\n"
        f"**Company**: {top_question['company']}\n"
        f"**Question Name**: {top_question['questionName']}\n"
        f"**Difficulty Level**: {top_question['difficulty level']}\n"
        f"**Tags**: {top_question['Tags']}\n"
        f"**Content**: {top_question['Content']}\n"
        f"\nPlease create a real-world interview question based on this information."
    )
    
    # Generate response using GPT-4 with detailed prompt and debugging logs
    response = generate_response([{"role": "assistant", "content": question_generation_prompt}, {"role": "user", "content": detailed_prompt}])

    # Display assistant response in chat message container and add to session history
    with st.chat_message("assistant"):
        st.markdown(response)
    
    st.session_state.messages.append({"role": "assistant", "content": response})
    
    # Enable follow-up mode after generating the initial question
    st.session_state.follow_up_mode = True

# Display chat messages from history on app rerun (for subsequent conversation)
for message in st.session_state.messages:
    with st.chat_message(message["role"]):
        st.markdown(message["content"])

# Chatbox for subsequent conversations with assistant (follow-up mode)
if st.session_state.follow_up_mode:
    if user_input := st.chat_input("Continue your conversation or ask follow-up questions here:"):
        # Display user message in chat message container and add to session history
        with st.chat_message("user"):
            st.markdown(user_input)
        
        st.session_state.messages.append({"role": "user", "content": user_input})

        # Generate assistant's response based on follow-up input using technical_interviewer_prompt as system prompt
        assistant_response = generate_response(
            [{"role": "assistant", "content": technical_interviewer_prompt}] + st.session_state.messages
        )

        with st.chat_message("assistant"):
            st.markdown(assistant_response)
        
        st.session_state.messages.append({"role": "assistant", "content": assistant_response})

st.sidebar.markdown("""
## About
This is a Real-World Interview Question Generator powered by OpenAI's API.
Enter a company name, topic, and level of difficulty, and it will transform a relevant question into a real-world interview scenario!
Continue chatting with the assistant in the chatbox below.
""")