app.py

import os
import nest_asyncio
nest_asyncio.apply()
import streamlit as st
from transformers import pipeline, AutoTokenizer
from huggingface_hub import login
from streamlit.components.v1 import html
import pandas as pd
import torch
import random

# Retrieve the token from environment variables
hf_token = os.environ.get("HF_TOKEN")
if not hf_token:
    st.error("Hugging Face token not found. Please set the HF_TOKEN environment variable.")
    st.stop()

# Login with the token
login(token=hf_token)

# Initialize session state for timer
#if 'timer_started' not in st.session_state:
    #st.session_state.timer_started = False
#if 'timer_frozen' not in st.session_state:
    #st.session_state.timer_frozen = False

# Timer component using HTML and JavaScript
def timer():
    return """
    <div id="timer" style="font-size:16px;color:#666;margin-bottom:10px;">⏱️ Elapsed: 00:00</div>
    <script>
    (function() {
        var start = Date.now();
        var timerElement = document.getElementById('timer');
        localStorage.removeItem("freezeTimer");
        var interval = setInterval(function() {
            if(localStorage.getItem("freezeTimer") === "true"){
                clearInterval(interval);
                timerElement.style.color = '#00cc00';
                return;
            }
            var elapsed = Date.now() - start;
            var minutes = Math.floor(elapsed / 60000);
            var seconds = Math.floor((elapsed % 60000) / 1000);
            timerElement.innerHTML = '⏱️ Elapsed: ' +
            (minutes < 10 ? '0' : '') + minutes + ':' +
            (seconds < 10 ? '0' : '') + seconds;
        }, 1000);
    })();
    </script>
    """

st.set_page_config(page_title="Review Scorer & Report Generator", page_icon="📝")
st.header("Review Scorer & Report Generator")

# Concise introduction
st.write("This model will score your reviews in your CSV file and generate a report based on your query and those results.")

# Load models with caching to avoid reloading on every run
@st.cache_resource
def load_models():
    score_pipe = None
    gemma_pipe = None
    
    try:
        st.info("Loading sentiment analysis model...")
        score_pipe = pipeline("text-classification", 
                              model="cardiffnlp/twitter-roberta-base-sentiment-latest", 
                              device=0 if torch.cuda.is_available() else -1)
        st.success("Sentiment analysis model loaded successfully!")
    except Exception as e:
        st.error(f"Error loading score model: {e}")
    
    try:
        st.info("Loading Gemma model...")
        # Load the tokenizer separately with the chat template
        tokenizer = AutoTokenizer.from_pretrained("google/gemma-3-1b-it")
        gemma_pipe = pipeline("text-generation", 
                              model="google/gemma-3-1b-it",
                              tokenizer=tokenizer,  # Pass the loaded tokenizer here
                              device=0, 
                              torch_dtype=torch.bfloat16)
        st.success("Gemma model loaded successfully!")
    except Exception as e:
        st.error(f"Error loading Gemma model: {e}")
        st.error(f"Detailed error: {type(e).__name__}: {str(e)}")
    
    return score_pipe, gemma_pipe

def extract_assistant_content(raw_response):
    """Extract only the assistant's content from the Gemma-3 response."""
    # Convert to string and work with it directly
    response_str = str(raw_response)

    # Look for the assistant's content marker
    assistant_marker = "'role': 'assistant', 'content': '"
    if assistant_marker in response_str:
        start_idx = response_str.find(assistant_marker) + len(assistant_marker)
        # Extract everything after the marker until the end or closing quote
        content = response_str[start_idx:]

        # Find the end of the content (last single quote before the end of the string or before closing curly brace)
        end_markers = ["'}", "'}]"]
        end_idx = len(content)
        for marker in end_markers:
            pos = content.rfind(marker)
            if pos != -1 and pos < end_idx:
                end_idx = pos

        return content[:end_idx]

    # Fallback - return the original response
    return response_str

score_pipe, gemma_pipe = load_models()


# Input: Query text for scoring and CSV file upload for candidate reviews
query_input = st.text_area("Enter your query text for analysis (this does not need to be part of the CSV):")
uploaded_file = st.file_uploader("Upload Reviews CSV File (must contain a 'reviewText' column)", type=["csv"])

if score_pipe is None or gemma_pipe is None:
    st.error("Model loading failed. Please check your model names, token permissions, and GPU configuration.")
else:
    candidate_docs = []
    if uploaded_file is not None:
        try:
            df = pd.read_csv(uploaded_file)
            if 'reviewText' not in df.columns:
                st.error("CSV must contain a 'reviewText' column.")
            else:
                candidate_docs = df['reviewText'].dropna().astype(str).tolist()
        except Exception as e:
            st.error(f"Error reading CSV file: {e}")
    
    if st.button("Generate Report"):
        # Reset timer state so that the timer always shows up
        st.session_state.timer_started = False
        st.session_state.timer_frozen = False
        if uploaded_file is None:
            st.error("Please upload a CSV file.")
        elif not candidate_docs:
            st.error("CSV must contain a 'reviewText' column.")
        elif not query_input.strip():
            st.error("Please enter a query text!")
        else:
            if not st.session_state.timer_started and not st.session_state.timer_frozen:
                st.session_state.timer_started = True
                html(timer(), height=50)
            status_text = st.empty()
            progress_bar = st.progress(0)
            
            # Stage 1: Score candidate documents using the provided query.
            status_text.markdown("**🔍 Scoring candidate documents...**")
            
           # Process each review individually to avoid memory issues
            scored_results = []
            for i, doc in enumerate(candidate_docs):
                # Update progress based on current document
                progress = int((i / len(candidate_docs)) * 50)  # First half of progress bar (0-50%)
                progress_bar.progress(progress)
                
                # Process single document with truncation to avoid tensor size mismatch
                try:
                    # Use the tokenizer to properly truncate the input
                    tokenizer = score_pipe.tokenizer
                    max_length = tokenizer.model_max_length  # Usually 512 for RoBERTa
                    
                    # Truncate the text using the tokenizer to ensure it fits
                    encoded_input = tokenizer(doc, truncation=True, max_length=max_length, return_tensors="pt")
                    # Decode back to text to get the truncated version
                    truncated_doc = tokenizer.decode(encoded_input["input_ids"][0], skip_special_tokens=True)
                    
                    # Now process the truncated document
                    result = score_pipe(truncated_doc)
                    scored_results.append(result[0])  # Get the first result
                except Exception as e:
                    st.warning(f"Error processing document {i}: {str(e)}")
                    # Add a placeholder result to maintain indexing
                    scored_results.append({"label": "ERROR", "score": 0})
                
                # Display occasional status updates for large datasets
                if i % max(1, len(candidate_docs) // 10) == 0:
                    status_text.markdown(f"**🔍 Scoring documents... ({i}/{len(candidate_docs)})**")
                    
            # Pair each review with its score assuming the output order matches the input order.
            scored_docs = list(zip(candidate_docs, [result["score"] for result in scored_results]))
            
            progress_bar.progress(67)
            
            # Stage 2: Generate Report using Gemma in the new messages format.
            status_text.markdown("**📝 Generating report with Gemma...**")

            # For very large datasets, summarize or sample the scored_docs before sending to Gemma
            sampled_docs = scored_docs
            if len(scored_docs) > 10000:  # Arbitrary threshold for what's "too large"
                # Option 1: Random sampling
                sampled_docs = random.sample(scored_docs, 1000)
            
            # Build the user content with query, sentiment results, and original review data.
            # Format the prompt as chat for Gemma
            messages = [
                {"role": "user", "content": f"""
Generate a concise 300-word report based on the following analysis without repeating what's in the analysis.
Query:
"{query_input}"
Candidate Reviews with their scores:
{scored_docs}
        """}
            ]
            
            raw_result = gemma_pipe(messages, max_new_tokens=50)
            report = extract_assistant_content(raw_result)
            progress_bar.progress(100)
            status_text.success("**✅ Generation complete!**")
            html("<script>localStorage.setItem('freezeTimer', 'true');</script>", height=0)
            st.session_state.timer_frozen = True
            #st.write("**Scored Candidate Reviews:**", scored_docs)
            st.write("**Generated Report:**", report)