app.py

import os
import numpy as np
import pandas as pd
import streamlit as st
from transformers import pipeline, AutoTokenizer, 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")

# ─── FLAN-T5 Generation Pipeline ────────────────────────────────────────────
@st.cache_resource
def load_flant5_pipeline():
    # High-level helper for text2text generation
    return pipeline(
        "text2text-generation",
        model="google/flan-t5-base",
        tokenizer="google/flan-t5-base"
    )

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

    progress = st.progress(0)

    # Load models
    progress.text("Loading models...")
    sentiment_pipeline = load_sentiment_pipeline()
    kw_model = load_keybert_model()
    generation_pipeline = load_flant5_pipeline()
    progress.progress(20)

    # Sentiment
    progress.text("Analyzing sentiment...")
    raw_scores = sentiment_pipeline(review)[0]
    sentiment_results = {LABEL_MAP[item['label']]: float(item['score']) for item in raw_scores}
    progress.progress(40)

    # Keywords
    progress.text("Extracting keywords...")
    keywords = kw_model.extract_keywords(
        review,
        keyphrase_ngram_range=(1, 2),
        stop_words="english",
        top_n=3
    )
    progress.progress(60)

    # Display
    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})")

    # Chart
    progress.text("Rendering chart...")
    df_scores = pd.DataFrame.from_dict(sentiment_results, orient='index', columns=['score'])
    df_scores.index.name = 'Sentiment'
    st.bar_chart(df_scores)
    progress.progress(80)

    # 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})")

    # FLAN-T5 Analysis & Suggestions
    progress.text("Generating insights...")
    prompt = f"""
You are an analytical amazon feedback expert.
Review: \"{review}\"
Sentiment Scores: {sentiment_results}
Top Keywords: {[kw for kw, _ in keywords]}

Tasks:
1. Analysis: Write a concise paragraph (3 sentences) interpreting customer sentiment by combining the scores and keywords.
2. Recommendations: Three separate paragraphs with actionable suggestions (max 30 words each).
"""
    output = generation_pipeline(prompt, max_length=200, do_sample=False)[0]['generated_text']
    st.markdown(output)

    progress.progress(100)
    progress.text("Done!")


if __name__ == "__main__":
    main()