app.py

import pandas as pd
import gradio as gr
from langchain.vectorstores import FAISS
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.schema import Document
import os
import pickle

EMBEDDING_MODEL = "sentence-transformers/multi-qa-mpnet-base-dot-v1"
DATASET_PATH = "qa_dataset.csv"
FAISS_INDEX_PATH = "faiss_index"

# Initialize system (Load dataset and FAISS index)
def initialize_system():
    if os.path.exists(FAISS_INDEX_PATH):
        print("Loading FAISS index from cache...")
        with open(FAISS_INDEX_PATH, "rb") as f:
            return pickle.load(f)

    print("Initializing FAISS from scratch...")
    data = pd.read_csv(DATASET_PATH).dropna().head(500)  # Limit rows for speed

    # Create documents
    documents = [
        Document(
            page_content=f"Q: {row['Question']}\nA: {row['Answer']}",
            metadata={"question": row['Question'], "answer": row['Answer']}
        ) for _, row in data.iterrows()
    ]

    # Load embedding model once
    embeddings = HuggingFaceEmbeddings(model_name=EMBEDDING_MODEL)

    # Create FAISS vector store
    vector_store = FAISS.from_documents(documents, embeddings)

    # Cache the FAISS index
    with open(FAISS_INDEX_PATH, "wb") as f:
        pickle.dump(vector_store, f)

    return vector_store

vector_store = initialize_system()

def classify_question(query: str, k: int = 3):
    results = vector_store.similarity_search(query, k=k)

    if not results:
        return {"Category": "Unknown", "Top Matches": "No matches found", "Confidence": "0%"}

    answers = " ".join([doc.metadata['answer'] for doc in results])
    keywords = list(dict.fromkeys(answers.split()))[:5]  # Extract first 5 unique words
    category = " ".join(keywords)

    return {
        "Category": category,
        "Top Matches": "\n\n".join([f"Q: {doc.metadata['question']}\nA: {doc.metadata['answer']}" 
                                    for doc in results]),
        "Confidence": f"{len(results)/k:.0%}"
    }

# Gradio interface
interface = gr.Interface(
    fn=lambda q: classify_question(q, 3),
    inputs=gr.Textbox(label="Input Question", placeholder="Type your question here..."),
    outputs=[
        gr.Textbox(label="Predicted Category"),
        gr.Textbox(label="Supporting Q&A"),
        gr.Textbox(label="Confidence")
    ],
    title="Question Classification System",
    description="Classify questions based on existing Q&A pairs using FAISS"
)

if __name__ == "__main__":
    interface.launch(share=True)