app.py

import streamlit as st
import os
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

# --- LANGCHAIN IMPORTS ---
from langchain_community.document_loaders import PyPDFLoader
from langchain_experimental.text_splitter import SemanticChunker
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_community.vectorstores import FAISS
from langchain.memory import ConversationBufferMemory

# 1) SET UP PAGE
st.title("💬 المحادثة التفاعلية - إدارة البيانات وحماية البيانات الشخصية")
local_file = "Policies001.pdf"

index_folder = "faiss_index"

# Inject custom CSS for right-to-left text
st.markdown(
    """
    <style>
        .rtl {
            direction: rtl;
            text-align: right;
        }
    </style>
    """,
    unsafe_allow_html=True
)

# 2) LOAD OR BUILD VECTORSTORE
embeddings = HuggingFaceEmbeddings(
    model_name="CAMeL-Lab/bert-base-arabic-camelbert-mix",
    model_kwargs={"trust_remote_code": True}
)

if os.path.exists(index_folder):
    vectorstore = FAISS.load_local(index_folder, embeddings, allow_dangerous_deserialization=True)
else:
    loader = PyPDFLoader(local_file)
    documents = loader.load()

    text_splitter = SemanticChunker(
        embeddings=embeddings,
        breakpoint_threshold_type='percentile',
        breakpoint_threshold_amount=90
    )
    chunked_docs = text_splitter.split_documents(documents)

    vectorstore = FAISS.from_documents(chunked_docs, embeddings)
    vectorstore.save_local(index_folder)

# 3) CREATE RETRIEVER
retriever = vectorstore.as_retriever(search_type="similarity", search_kwargs={"k": 5})

# 4) SET UP "COMMAND-R7B-ARABIC" AS LLM
# Authenticate and load the model
model_name = "CohereForAI/c4ai-command-r7b-arabic-02-2025"  # Replace with the actual Hugging Face model ID

# Set Hugging Face token securely
hf_token = os.getenv("HF_TOKEN")  # Ensure you set your token as an environment variable in Hugging Face Spaces

if hf_token is None:
    st.error("Hugging Face token not found. Please set the 'HF_TOKEN' environment variable.")
    st.stop()

# Load tokenizer and model using the token
tokenizer = AutoTokenizer.from_pretrained(model_name, use_auth_token=hf_token)
model = AutoModelForCausalLM.from_pretrained(model_name, use_auth_token=hf_token)

# Hugging Face pipeline for text generation
qa_pipeline = pipeline("text-generation", model=model, tokenizer=tokenizer, device=0)

# Memory object to store conversation
memory = ConversationBufferMemory(
    memory_key="chat_history",  # key used internally by the chain
    return_messages=True        # ensures we get the entire message history
)

# 5) MANAGE SESSION STATE FOR UI CHAT
if "messages" not in st.session_state:
    st.session_state["messages"] = [
        {"role": "assistant", "content": "👋 مرحبًا! اسألني أي شيء عن إدارة البيانات وحماية البيانات الشخصية!"}
    ]

# Display existing messages in chat format
for msg in st.session_state["messages"]:
    with st.chat_message(msg["role"]):
        # Apply the "rtl" class to style Arabic text correctly
        st.markdown(f'<div class="rtl">{msg["content"]}</div>', unsafe_allow_html=True)

# 6) CHAT INPUT
user_input = st.chat_input("اكتب سؤالك هنا")

# 7) PROCESS NEW USER MESSAGE
if user_input:
    # a) Display user message in UI
    st.session_state["messages"].append({"role": "user", "content": user_input})
    with st.chat_message("user"):
        st.markdown(f'<div class="rtl">{user_input}</div>', unsafe_allow_html=True)

    # b) Run pipeline to generate a response
    # Combine retriever results and user input for context-aware answering
    retrieved_docs = retriever.get_relevant_documents(user_input)
    context = "\n".join([doc.page_content for doc in retrieved_docs])
    full_input = f"السياق:\n{context}\n\nالسؤال:\n{user_input}"

    # Generate answer using the pipeline
    response = qa_pipeline(full_input, max_length=500, num_return_sequences=1)[0]["generated_text"]

    # c) Display assistant response
    st.session_state["messages"].append({"role": "assistant", "content": response})
    with st.chat_message("assistant"):
        st.markdown(f'<div class="rtl">{response}</div>', unsafe_allow_html=True)