app.py

import os
import numpy as np
import pandas as pd
import streamlit as st
from transformers import (
    pipeline,
    AutoTokenizer,
    AutoModelForSequenceClassification,
    AutoModelForSeq2SeqLM
)
from keybert import KeyBERT

# ─── Sentiment & Keyword Models ─────────────────────────────────────────────
@st.cache_resource
def load_sentiment_pipeline():
    model_name = "mayf/amazon_reviews_bert_ft"
    tok = AutoTokenizer.from_pretrained(model_name, use_auth_token=True)
    mdl = AutoModelForSequenceClassification.from_pretrained(
        model_name,
        use_auth_token=True
    )
    return pipeline(
        "sentiment-analysis",
        model=mdl,
        tokenizer=tok,
        return_all_scores=True
    )

@st.cache_resource
def load_keybert_model():
    return KeyBERT(model="all-MiniLM-L6-v2")

# ─── BlenderBot Response Components ─────────────────────────────────────────
@st.cache_resource
def load_response_components():
    # Load tokenizer and model directly for manual generation with truncation
    tok = AutoTokenizer.from_pretrained(
        "facebook/blenderbot-400M-distill",
        use_fast=True
    )
    mdl = AutoModelForSeq2SeqLM.from_pretrained("facebook/blenderbot-400M-distill")
    return tok, mdl
:
    # Use BlenderBot 400M Distill for text generation
    return pipeline(
        "text2text-generation",
        model="facebook/blenderbot-400M-distill",
        tokenizer="facebook/blenderbot-400M-distill",
        max_new_tokens=150,
        do_sample=False
    )

LABEL_MAP = {
    "LABEL_0": "Very Negative",
    "LABEL_1": "Negative",
    "LABEL_2": "Neutral",
    "LABEL_3": "Positive",
    "LABEL_4": "Very Positive"
}


def main():
    st.title("📊 Amazon Review Analyzer")

    review = st.text_area("Enter your review:")
    if not st.button("Analyze Review"):
        return
    if not review:
        st.warning("Please enter a review to analyze.")
        return

    # ─── KEEP THIS BLOCK UNCHANGED ─────────────────────────────────────────
    # Sentiment Analysis
    sentiment_pipeline = load_sentiment_pipeline()
    raw_scores = sentiment_pipeline(review)[0]
    sentiment_results = {LABEL_MAP[item['label']]: float(item['score']) for item in raw_scores}

    # Keyword Extraction
    kw_model = load_keybert_model()
    keywords = kw_model.extract_keywords(
        review,
        keyphrase_ngram_range=(1, 2),
        stop_words="english",
        top_n=3
    )

    # Display Results
    col1, col2 = st.columns(2)
    with col1:
        st.subheader("Sentiment Scores")
        st.json({k: round(v, 4) for k, v in sentiment_results.items()})
    with col2:
        st.subheader("Top 3 Keywords")
        for kw, score in keywords:
            st.write(f"• {kw} ({score:.4f})")

    # Bar Chart
    df_scores = pd.DataFrame.from_dict(
        sentiment_results,
        orient='index',
        columns=['score']
    )
    df_scores.index.name = 'Sentiment'
    st.bar_chart(df_scores)

    # Highlight Highest Sentiment
    max_label, max_score = max(sentiment_results.items(), key=lambda x: x[1])
    st.markdown(f"**Highest Sentiment:** **{max_label}** ({max_score:.4f})")
    # ────────────────────────────────────────────────────────────────────

    # Generate appropriate reply using manual tokenization & generation
    tok, mdl = load_response_components()
    if max_label in ["Positive", "Very Positive"]:
        prompt_text = (
            f"You are a friendly customer success representative. The customer said: \"{review}\". "
            "Write two sentences to express gratitude and highlight their positive experience."
        )
    else:
        prompt_text = (
            f"You are a helpful customer support specialist. The customer said: \"{review}\". "
            f"Identified issues: {', '.join([kw for kw, _ in keywords])}. "
            "First, ask 1-2 clarifying questions to understand their situation. "
            "Then provide two concrete suggestions or next steps to address these issues."
        )
    # Tokenize with truncation to avoid out-of-range embeddings
    inputs = tok(
        prompt_text,
        return_tensors="pt",
        truncation=True,
        max_length=tok.model_max_length
    )
    outputs = mdl.generate(
        **inputs,
        max_new_tokens=150,
        do_sample=False
    )
    reply = tok.decode(outputs[0], skip_special_tokens=True).strip()

    st.subheader("Generated Reply")
    st.write(reply)


if __name__ == '__main__':
    main()