app.py

import gradio as gr
import pickle
from rank_bm25 import BM25Okapi
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity
from difflib import SequenceMatcher
import numpy as np
import random

# Load model
print("🤖 Loading model...")
with open('chatbot_caca.pkl', 'rb') as f:
    data = pickle.load(f)

qa_pairs = data['qa_pairs']
bm25 = data['bm25']
tfidf = data['tfidf']
tfidf_matrix = data['tfidf_matrix']
answers = data['answers']

print(f"✅ Loaded {len(qa_pairs)} QA pairs")


def preprocess(text):
    return text.lower().strip()


def get_bm25_score(user_input, top_k=3):
    tokenized_query = preprocess(user_input).split()
    scores = bm25.get_scores(tokenized_query)
    top_indices = np.argsort(scores)[-top_k:][::-1]
    return [(idx, scores[idx]) for idx in top_indices]


def get_tfidf_score(user_input, top_k=3):
    user_vector = tfidf.transform([preprocess(user_input)])
    similarities = cosine_similarity(user_vector, tfidf_matrix)[0]
    top_indices = np.argsort(similarities)[-top_k:][::-1]
    return [(idx, similarities[idx]) for idx in top_indices]


def get_fuzzy_score(user_input, candidate_idx):
    question = qa_pairs[candidate_idx]['question']
    return SequenceMatcher(None, preprocess(user_input), preprocess(question)).ratio()


def fallback_response(confidence=0.0):
    if confidence > 0.15:
        responses = [
            "hmm kayaknya aku tau sih maksudmu, tapi ga terlalu yakin... coba tanya dengan kata lain? 🤔",
            "aku nangkep sedikit sih, tapi ga confident buat jawab. bisa diperjelas ga?",
        ]
    else:
        responses = [
            "waduh, pertanyaan ini di luar kemampuanku nih. Lyon-nya kurang ngajarin kayaknya 🙄",
            "jujur aja ya, aku ga ngerti maksudmu 😂 coba tanya yang lain deh",
            "kayaknya pertanyaan ini terlalu advanced buat AI bernama Caca Kecil 😅",
            "hmm aku belum tau jawabannya nih. Lyon-nya lagi males update dataset kayaknya 😤",
            "maaf belum bisa jawab yang itu. tapi aku usahain belajar ya! *semangat meski nama ngaco*",
        ]
    return random.choice(responses)


def chat(message, history):
    """Chat function untuk Gradio"""
    
    # Get scores
    bm25_results = get_bm25_score(message, top_k=3)
    tfidf_results = get_tfidf_score(message, top_k=3)
    
    # Combine scores
    combined_scores = {}
    
    for idx, score in bm25_results:
        normalized_score = min(score / 20, 1.0)
        combined_scores[idx] = combined_scores.get(idx, 0) + (normalized_score * 0.4)
    
    for idx, score in tfidf_results:
        combined_scores[idx] = combined_scores.get(idx, 0) + (score * 0.5)
    
    if not combined_scores:
        return fallback_response(0.0)
    
    best_idx = max(combined_scores, key=combined_scores.get)
    best_score = combined_scores[best_idx]
    
    # Fuzzy bonus
    fuzzy_score = get_fuzzy_score(message, best_idx)
    final_score = best_score + (fuzzy_score * 0.1)
    
    threshold = 0.25
    
    if final_score >= threshold:
        return answers[best_idx]
    else:
        return fallback_response(final_score)


# Create Gradio interface
demo = gr.ChatInterface(
    fn=chat,
    title="💬 Chatbot Caca",
    description="""
    Chatbot berbasis retrieval (BM25 + TF-IDF) untuk QA Bahasa Indonesia.
    
    **Fun fact:** AI ini namanya Caca Kecil karena creator-nya (Lyon) punya selera penamaan yang... unik 😂
    
    Model size: 2.83 MB | QA pairs: 3,500+ | No LLM needed!
    """,
    examples=[
        "siapa nama kamu?",
        "ceritakan tentang dirimu",
        "siapa itu Lyon?",
        "kenapa namamu Caca?",
        "kamu bisa apa?",
    ],
    theme="soft",
    chatbot=gr.Chatbot(height=400),
)

if __name__ == "__main__":
    demo.launch()