model.py

import os
os.environ["TRANSFORMERS_CACHE"] = "/tmp/hf-cache"
os.environ["HF_HOME"] = "/tmp/hf-home"
import nltk
os.environ["NLTK_DATA"] = "/tmp/nltk_data"
nltk.download("punkt", download_dir="/tmp/nltk_data")  
from typing import List, Optional
from pydantic import BaseModel
from transformers import pipeline

# ✅ Extra: Smart Summarization Imports
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.cluster import KMeans
from nltk.tokenize import sent_tokenize
from sklearn.metrics.pairwise import cosine_similarity
import numpy as np

summarizer = pipeline("summarization", model="sshleifer/distilbart-cnn-12-6")
sentiment_analyzer = pipeline("sentiment-analysis")

# 🧠 Basic Summarization (Abstractive)
def summarize_review(text):
    return summarizer(text, max_length=60, min_length=10, do_sample=False, no_repeat_ngram_size=3)[0]["summary_text"]

# 🧠 Smart Summarization (Clustered Key Sentences)
def smart_summarize(text, n_clusters=1):
    """Improved summarization using clustering on sentence embeddings"""
    tokenizer = nltk.tokenize.PunktSentenceTokenizer()  # ✅ Use default trained Punkt tokenizer
    sentences = tokenizer.tokenize(text)

    if len(sentences) <= 1:
        return text

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

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

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

    avg = []
    for i in range(n_clusters):
        idx = np.where(kmeans.labels_ == i)[0]
        if len(idx) == 0:
            continue
        avg_vector = tfidf_matrix[idx].mean(axis=0).A1.reshape(1, -1)  # Convert np.matrix to ndarray
        sim = cosine_similarity(avg_vector, tfidf_matrix[idx])
        most_representative_idx = idx[np.argmax(sim)]
        avg.append(sentences[most_representative_idx])

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

# 📊 Sentiment Detection
def analyze_sentiment(text):
    result = sentiment_analyzer(text)[0]
    label = result["label"]
    score = result["score"]

    if "star" in label:
        stars = int(label[0])
        if stars <= 2:
            label = "NEGATIVE"
        elif stars == 3:
            label = "NEUTRAL"
        else:
            label = "POSITIVE"

    return {
        "label": label,
        "score": score
    }

# 🔥 Emotion Detection (heuristic-based)
def detect_emotion(text):
    text_lower = text.lower()
    if "angry" in text_lower or "hate" in text_lower:
        return "anger"
    elif "happy" in text_lower or "love" in text_lower:
        return "joy"
    elif "sad" in text_lower or "disappointed" in text_lower:
        return "sadness"
    elif "confused" in text_lower or "unclear" in text_lower:
        return "confusion"
    else:
        return "neutral"

# 🧩 Aspect-Based Sentiment (mock)
def extract_aspect_sentiment(text, aspects: list):
    results = {}
    text_lower = text.lower()
    for asp in aspects:
        label = "positive" if asp in text_lower and "not" not in text_lower else "neutral"
        results[asp] = {
            "label": label,
            "confidence": 0.85
        }
    return results

# ✅ Pydantic Schemas for FastAPI
class ReviewInput(BaseModel):
    text: str
    model: str = "distilbert-base-uncased-finetuned-sst-2-english"
    industry: str = "Generic"
    aspects: bool = False
    follow_up: Optional[str] = None
    product_category: Optional[str] = None
    device: Optional[str] = None

class BulkReviewInput(BaseModel):
    reviews: List[str]
    model: str = "distilbert-base-uncased-finetuned-sst-2-english"
    industry: str = "Generic"
    aspects: bool = False
    product_category: Optional[str] = None
    device: Optional[str] = None

class TranslationInput(BaseModel):
    text: str
    target_lang: str = "fr"

class ChatInput(BaseModel):
    question: str
    context: str