src/streamlit_app.py

import streamlit as st
import torch
import os
import pickle
import torch.nn.functional as F
from transformers import BertTokenizer, BertForSequenceClassification, AutoTokenizer, AutoModelForSeq2SeqLM, AutoModelForSequenceClassification
import asyncio
import sys

if sys.platform == "win32":
    asyncio.set_event_loop_policy(asyncio.WindowsSelectorEventLoopPolicy())


# from huggingface_hub import login

# token = os.getenv("HF_TOKEN")  # Read from environment
# if token:
#     login(token=token)
# else:
#     raise ValueError("HF_TOKEN environment variable is not set.")
# -------------------- Part 1: Prediction --------------------

@st.cache_resource
def load_prediction_model():
    tokenizer = BertTokenizer.from_pretrained('https://huggingface.co/bert-base-uncased')
    with open('src/label_encoder.pkl', 'rb') as f:
        label_encoder = pickle.load(f)
    id_to_class = {idx: class_name for idx, class_name in enumerate(label_encoder.classes_)}

    model = BertForSequenceClassification.from_pretrained('Divyanshu04/Issue_categorizerv2')
    # model.load_state_dict(torch.load('Divyanshu04/Issue_categorizer', map_location=torch.device('cpu'))['model_state_dict'])
    # model.eval()
    return tokenizer, model, id_to_class

tokenizer_cls, model_cls, id_to_class = load_prediction_model()

def preprocess_texts(texts):
    return tokenizer_cls(texts, padding='max_length', truncation=True, max_length=128, return_tensors='pt')

def predict(text):
    inputs = preprocess_texts(text)
    with torch.no_grad():
        outputs = model_cls(**inputs)
        probabilities = F.softmax(outputs.logits, dim=1)
        top3_probs, top3_classes = torch.topk(probabilities, k=3, dim=1)

    top3_class_names = [id_to_class[idx.item()] for idx in top3_classes[0]]
    top3_probs = top3_probs[0] * 100
    top3_probs_np = top3_probs.cpu().numpy()
    formatted_percentages = [f"{p:.4f}%" for p in top3_probs_np]

    prediction = top3_class_names[0]
    probability = top3_probs_np[0]

    return top3_class_names, top3_probs_np, formatted_percentages, prediction, probability

# -------------------- Part 2: Follow-up Generator --------------------

@st.cache_resource
def load_followup_model():
    model_path = "https://huggingface.co/Divyanshu04/Insurance_claim_followup_model"  # Adjust path as needed
    tokenizer = AutoTokenizer.from_pretrained(model_path)
    model = AutoModelForSeq2SeqLM.from_pretrained(model_path)
    model.eval()
    return tokenizer, model

tokenizer_seq, model_seq = load_followup_model()

def generate_followup(context, condition=None, max_tokens=64):
    prompt = f"Context: {context}"
    if condition:
        prompt += f"\nCondition: {condition}"
    prompt += "\nFollow-up question:"
    inputs = tokenizer_seq(prompt, return_tensors="pt", padding=True).to(model_seq.device)

    outputs = model_seq.generate(
        **inputs,
        max_new_tokens=max_tokens,
        do_sample=True,
        top_k=50,
        top_p=0.95,
        temperature=0.9,
        num_return_sequences=1
    )
    return tokenizer_seq.decode(outputs[0], skip_special_tokens=True)

# -------------------- Streamlit UI --------------------

st.title("Tenant-Landlord Query Classifier + Claim Assistant")

if "prev_input" not in st.session_state:
    st.session_state.prev_input = ""

if "context" not in st.session_state:
    st.session_state.context = ""

user_input = st.text_input("Enter your query:")

if user_input:
    with st.spinner("Classifying your query..."):
        combined_input = st.session_state.prev_input + " " + user_input if st.session_state.prev_input else user_input
        top3_classes, top3_probs, formatted, prediction, probability = predict(combined_input)

        if probability > 60:
            st.success(f"Prediction: **{prediction}** with confidence **{probability:.2f}%**")
            st.write("Top 3 predictions:")
            for cls, prob in zip(top3_classes, formatted):
                st.write(f"- {cls}: {prob}")
            st.session_state.context = combined_input
            st.session_state.prev_input = ""  # Reset
        else:
            st.warning("Confidence is low. Please elaborate your query more.")
            st.session_state.prev_input = combined_input

# -------------------- Ask to Make a Claim --------------------

if st.session_state.context:
    make_claim = st.radio("Do you want to make a claim?", ["Yes", "No"])

    if make_claim == "No":
        st.info("Thank you! No claim will be made.")
        st.stop()

    elif make_claim == "Yes":
        st.subheader("Claim Assistant - Answer Follow-up Questions")

        if "followup_count" not in st.session_state:
            st.session_state.followup_count = 0
        if "questions" not in st.session_state:
            st.session_state.questions = []
        if "responses" not in st.session_state:
            st.session_state.responses = []

        # Generate new follow-up question if needed
        if len(st.session_state.questions) <= st.session_state.followup_count and st.session_state.followup_count < 5:
            with st.spinner("Generating follow-up question..."):
                new_question = generate_followup(st.session_state.context)
            st.session_state.questions.append(new_question)

        # Render follow-up questions and collect responses
        for i in range(len(st.session_state.questions)):
            st.markdown(f"**Follow-up Question {i+1}:** {st.session_state.questions[i]}")
            response_key = f"response_input_{i}"
            response = st.text_input(f"Your response to question {i+1}:", key=response_key)

            if response and len(st.session_state.responses) <= i:
                st.session_state.responses.append(response)
                st.session_state.context += " " + response
                st.session_state.followup_count += 1

        if st.session_state.followup_count >= 5:
            st.success("All follow-up questions answered. Your claim has been registered.")