app.py

import streamlit as st
from transformers import AutoTokenizer, AutoModelForCausalLM, TextIteratorStreamer
from huggingface_hub import login
from threading import Thread
import PyPDF2
import pandas as pd
import torch
import time
import os
from transformers import AutoModelForMaskedLM


# Check if 'peft' is installed
try:
    from peft import PeftModel, PeftConfig
except ImportError:
    raise ImportError(
        "The 'peft' library is required but not installed. "
        "Please install it using: `pip install peft`"
    )

# 🔐 Hardcoded Hugging Face Token

HF_TOKEN = os.environ.get("HF_TOKEN") # Replace with your actual token

# Set page configuration
st.set_page_config(
    page_title="Assistente LGT | Angola",
    page_icon="🚀",
    layout="centered"
)

# Model base and options
BASE_MODEL_NAME = "neuralmind/bert-base-portuguese-cased" #"pierreguillou/gpt2-small-portuguese" #"unicamp-dl/ptt5-base-portuguese-vocab" #"mistralai/Mistral-7B-Instruct-v0.2"
MODEL_OPTIONS = {
    "Full Fine-Tuned": "amiguel/mistral-angolan-laborlaw-bert-base-pt", #"amiguel/mistral-angolan-laborlaw-gpt2",#"amiguel/mistral-angolan-laborlaw-ptt5", #"amiguel/mistral-angolan-laborlaw",
    "LoRA Adapter": "amiguel/SmolLM2-360M-concise-reasoning-lora",
    "QLoRA Adapter": "amiguel/SmolLM2-360M-concise-reasoning-qlora"
}

st.title("🚀 Assistente LGT | Angola 🚀")

USER_AVATAR = "https://raw.githubusercontent.com/achilela/vila_fofoka_analysis/9904d9a0d445ab0488cf7395cb863cce7621d897/USER_AVATAR.png"
BOT_AVATAR = "https://raw.githubusercontent.com/achilela/vila_fofoka_analysis/991f4c6e4e1dc7a8e24876ca5aae5228bcdb4dba/Ataliba_Avatar.jpg"

# Sidebar
with st.sidebar:
    st.header("Model Selection 🤖")
    model_type = st.selectbox("Choose Model Type", list(MODEL_OPTIONS.keys()), index=0)
    selected_model = MODEL_OPTIONS[model_type]

    st.header("Upload Documents 📂")
    uploaded_file = st.file_uploader(
        "Choose a PDF or XLSX file",
        type=["pdf", "xlsx"],
        label_visibility="collapsed"
    )

# Session state
if "messages" not in st.session_state:
    st.session_state.messages = []

# File processor
@st.cache_data
def process_file(uploaded_file):
    if uploaded_file is None:
        return ""
    
    try:
        if uploaded_file.type == "application/pdf":
            pdf_reader = PyPDF2.PdfReader(uploaded_file)
            return "\n".join([page.extract_text() for page in pdf_reader.pages])
        elif uploaded_file.type == "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet":
            df = pd.read_excel(uploaded_file)
            return df.to_markdown()
    except Exception as e:
        st.error(f"📄 Error processing file: {str(e)}")
        return ""

# Model loader
@st.cache_resource
def load_model(model_type, selected_model):
    try:
        login(token=HF_TOKEN)

        tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_NAME, token=HF_TOKEN)

        if model_type == "Full Fine-Tuned":
        
             model = AutoModelForMaskedLM.from_pretrained(
                 selected_model,
                 torch_dtype=torch.bfloat16,  # or float32 for compatibility
                 token=HF_TOKEN
            ).to("cuda" if torch.cuda.is_available() else "cpu")

            #model = AutoModelForCausalLM.from_pretrained(
                 #selected_model,
       #         torch_dtype=torch.bfloat16,
       #         device_map="auto",
       #         token=HF_TOKEN
       #     
        
        else:
            base_model = AutoModelForCausalLM.from_pretrained(
                BASE_MODEL_NAME,
                torch_dtype=torch.bfloat16,
                device_map="auto",
                token=HF_TOKEN
            )
            model = PeftModel.from_pretrained(
                base_model,
                selected_model,
                torch_dtype=torch.bfloat16,
                is_trainable=False,
                token=HF_TOKEN
            )
        return model, tokenizer

    except Exception as e:
        st.error(f"🤖 Model loading failed: {str(e)}")
        return None

# Generation function
def generate_with_kv_cache(prompt, file_context, model, tokenizer, use_cache=True):
    full_prompt = f"Analyze this context:\n{file_context}\n\nQuestion: {prompt}\nAnswer:"
    
    streamer = TextIteratorStreamer(
        tokenizer, 
        skip_prompt=True, 
        skip_special_tokens=True
    )
    
    inputs = tokenizer(full_prompt, return_tensors="pt").to(model.device)
    
    generation_kwargs = {
        "input_ids": inputs["input_ids"],
        "attention_mask": inputs["attention_mask"],
        "max_new_tokens": 1024,
        "temperature": 0.7,
        "top_p": 0.9,
        "repetition_penalty": 1.1,
        "do_sample": True,
        "use_cache": use_cache,
        "streamer": streamer
    }
    
    Thread(target=model.generate, kwargs=generation_kwargs).start()
    return streamer

# Display chat history
for message in st.session_state.messages:
    avatar = USER_AVATAR if message["role"] == "user" else BOT_AVATAR
    with st.chat_message(message["role"], avatar=avatar):
        st.markdown(message["content"])

# Prompt interaction
if prompt := st.chat_input("Ask your inspection question..."):

    # Load model if necessary
    if "model" not in st.session_state or st.session_state.get("model_type") != model_type:
        model_data = load_model(model_type, selected_model)
        if model_data is None:
            st.error("Failed to load model.")
            st.stop()

        st.session_state.model, st.session_state.tokenizer = model_data
        st.session_state.model_type = model_type

    model = st.session_state.model
    tokenizer = st.session_state.tokenizer

    with st.chat_message("user", avatar=USER_AVATAR):
        st.markdown(prompt)
    st.session_state.messages.append({"role": "user", "content": prompt})

    file_context = process_file(uploaded_file)

    if model and tokenizer:
        try:
            with st.chat_message("assistant", avatar=BOT_AVATAR):
                start_time = time.time()
                streamer = generate_with_kv_cache(prompt, file_context, model, tokenizer, use_cache=True)

                response_container = st.empty()
                full_response = ""

                for chunk in streamer:
                    cleaned_chunk = chunk.replace("<think>", "").replace("</think>", "").strip()
                    full_response += cleaned_chunk + " "
                    response_container.markdown(full_response + "▌", unsafe_allow_html=True)

                end_time = time.time()
                input_tokens = len(tokenizer(prompt)["input_ids"])
                output_tokens = len(tokenizer(full_response)["input_ids"])
                speed = output_tokens / (end_time - start_time)

                input_cost = (input_tokens / 1_000_000) * 5
                output_cost = (output_tokens / 1_000_000) * 15
                total_cost_usd = input_cost + output_cost
                total_cost_aoa = total_cost_usd * 1160

                st.caption(
                    f"🔑 Input Tokens: {input_tokens} | Output Tokens: {output_tokens} | "
                    f"🕒 Speed: {speed:.1f}t/s | 💰 Cost (USD): ${total_cost_usd:.4f} | "
                    f"💵 Cost (AOA): {total_cost_aoa:.4f}"
                )

                response_container.markdown(full_response)
                st.session_state.messages.append({"role": "assistant", "content": full_response})

        except Exception as e:
            st.error(f"⚡ Generation error: {str(e)}")
    else:
        st.error("🤖 Model not loaded!")