import os
os.environ["TRANSFORMERS_CACHE"] = "/tmp/hf-cache"
os.environ["HF_HOME"] = "/tmp/hf-home"

import nltk
nltk.download("punkt", download_dir="/tmp/nltk_data")

from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.cluster import KMeans
from sklearn.metrics.pairwise import cosine_similarity
from nltk.tokenize import sent_tokenize
from transformers import pipeline
import numpy as np

# Load summarizer and Q&A pipeline
summarizer = pipeline("summarization", model="sshleifer/distilbart-cnn-12-6")
qa_pipeline = pipeline("question-answering", model="distilbert-base-cased-distilled-squad")

# --- Summarization Functions ---
def summarize_review(text):
    """Standard transformer-based summarization"""
    return summarizer(text, max_length=max_len, min_length=min_len, do_sample=False)[0]["summary_text"]
def smart_summarize(text, n_clusters=1):
    """
    Clustering + cosine similarity-based summarization
    Selects most representative sentence(s) from each cluster
    """
    sentences = sent_tokenize(text)
    if len(sentences) <= 1:
        return text

    tfidf = TfidfVectorizer(stop_words="english")
    tfidf_matrix = tfidf.fit_transform(sentences)

    if len(sentences) <= n_clusters:
        return " ".join(sentences)

    kmeans = KMeans(n_clusters=n_clusters, random_state=42).fit(tfidf_matrix)
    summary_sentences = []

    for i in range(n_clusters):
        idx = np.where(kmeans.labels_ == i)[0]
        if not len(idx):
            continue
        avg_vector = np.asarray(tfidf_matrix[idx].mean(axis=0))
        sim = cosine_similarity(avg_vector, tfidf_matrix[idx].toarray())
        most_representative = sentences[idx[np.argmax(sim)]]
        summary_sentences.append(most_representative)

    return " ".join(sorted(summary_sentences, key=sentences.index))

# --- Rule-based Category Detectors ---
def detect_industry(text):
    text = text.lower()
    if any(k in text for k in ["doctor", "hospital", "health", "pill", "med"]):
        return "Healthcare"
    if any(k in text for k in ["flight", "hotel", "trip", "booking"]):
        return "Travel"
    if any(k in text for k in ["bank", "loan", "credit", "payment"]):
        return "Banking"
    if any(k in text for k in ["gym", "trainer", "fitness", "workout"]):
        return "Fitness"
    if any(k in text for k in ["movie", "series", "stream", "video"]):
        return "Entertainment"
    if any(k in text for k in ["game", "gaming", "console"]):
        return "Gaming"
    if any(k in text for k in ["food", "delivery", "restaurant", "order"]):
        return "Food Delivery"
    if any(k in text for k in ["school", "university", "teacher", "course"]):
        return "Education"
    if any(k in text for k in ["insurance", "policy", "claim"]):
        return "Insurance"
    if any(k in text for k in ["property", "rent", "apartment", "house"]):
        return "Real Estate"
    if any(k in text for k in ["shop", "buy", "product", "phone", "amazon", "flipkart"]):
        return "E-commerce"
    return "Generic"

def detect_product_category(text):
    text = text.lower()
    if any(k in text for k in ["mobile", "smartphone", "iphone", "samsung", "phone"]):
        return "Mobile Devices"
    if any(k in text for k in ["laptop", "macbook", "notebook", "chromebook"]):
        return "Laptops"
    if any(k in text for k in ["tv", "refrigerator", "microwave", "washer"]):
        return "Home Appliances"
    if any(k in text for k in ["watch", "band", "fitbit", "wearable"]):
        return "Wearables"
    if any(k in text for k in ["app", "portal", "site", "website"]):
        return "Web App"
    return "General"

# --- Follow-up Q&A ---
def answer_followup(text, question, verbosity="brief"):
    try:
        response = qa_pipeline({"question": question, "context": text})
        answer = response.get("answer", "")
        if verbosity.lower() == "detailed":
            return f"Based on the review, the answer is: **{answer}**"
        return answer
    except Exception:
        return "Sorry, I couldn't generate a follow-up answer."