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")

# ─── FLAN-T5 Generation Pipeline ────────────────────────────────────────────
@st.cache_resource
def load_flant5_pipeline():
    # Explicitly load the Seq2Seq model & tokenizer to avoid truncation/classification fallback
    seq_tok = AutoTokenizer.from_pretrained("google/flan-t5-large")
    seq_model = AutoModelForSeq2SeqLM.from_pretrained("google/flan-t5-large")
    return pipeline(
        "text2text-generation",
        model=seq_model,
        tokenizer=seq_tok,
        # ensure we generate up to 400 new tokens
        max_new_tokens=400,
        do_sample=True,
        temperature=0.7
    )

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 Analysis
    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)

    # Keyword Extraction
    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 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
    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]}

Please complete the following:
- ANALYSIS: A concise paragraph (3 sentences) interpreting customer sentiment.
- RECOMMENDATIONS: Three separate paragraphs with actionable suggestions (max 30 words each).
"""
    response = generation_pipeline(prompt)
    output = response[0]["generated_text"]
    st.markdown(output)

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


if __name__ == "__main__":
    main()