app.py

# app.py

import gradio as gr
import os
import pdfminer.high_level
import docx
from sentence_transformers import SentenceTransformer
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
import faiss
import tempfile

# ====== Settings ======
EMBEDDING_MODEL_NAME = "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2"
GENERATION_MODEL_NAME = "aubmindlab/aragpt2-small"
CHUNK_SIZE = 500
CHUNK_OVERLAP = 50
TOP_K = 5

# ====== Load Models ======
embedder = SentenceTransformer(EMBEDDING_MODEL_NAME)
gen_tokenizer = AutoTokenizer.from_pretrained(GENERATION_MODEL_NAME)
gen_model = AutoModelForCausalLM.from_pretrained(GENERATION_MODEL_NAME)

# ====== Globals ======
index = None
chunks = []

# ====== Helpers ======
def extract_text_from_pdf(file_path):
    with open(file_path, 'rb') as f:
        return pdfminer.high_level.extract_text(f)

def extract_text_from_docx(file_path):
    doc = docx.Document(file_path)
    return "\n".join([para.text for para in doc.paragraphs])

def chunk_text(text):
    words = text.split()
    chunks = []
    for i in range(0, len(words), CHUNK_SIZE - CHUNK_OVERLAP):
        chunk = " ".join(words[i:i+CHUNK_SIZE])
        chunks.append(chunk)
    return chunks

def build_vector_store(chunks):
    vectors = embedder.encode(chunks)
    dim = vectors.shape[1]
    idx = faiss.IndexFlatL2(dim)
    idx.add(vectors)
    return idx, vectors

def retrieve_relevant_chunks(question, idx, chunks, vectors):
    q_vec = embedder.encode([question])
    D, I = idx.search(q_vec, TOP_K)
    return [chunks[i] for i in I[0] if i < len(chunks)]

def generate_answer(context_chunks, question):
    context = " \n".join(context_chunks)
    prompt = f"سؤال: {question}\nمحتوى ذو صلة: {context}\nجواب:"
    inputs = gen_tokenizer(prompt, return_tensors="pt", max_length=1024, truncation=True)
    outputs = gen_model.generate(**inputs, max_new_tokens=100)
    answer = gen_tokenizer.decode(outputs[0], skip_special_tokens=True)
    answer = answer.split("جواب:")[-1].strip()
    return answer

# ====== Gradio Functions ======
def upload_and_train(files):
    global index, chunks

    all_text = ""
    for file in files:
        suffix = os.path.splitext(file.name)[-1].lower()
        with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:
            tmp.write(file.read())
            tmp_path = tmp.name
        if suffix == ".pdf":
            all_text += extract_text_from_pdf(tmp_path) + "\n"
        elif suffix in [".docx", ".doc"]:
            all_text += extract_text_from_docx(tmp_path) + "\n"
        os.unlink(tmp_path)

    chunks = chunk_text(all_text)
    index, vectors = build_vector_store(chunks)

    return "✅ النظام جاهز للإجابة على أسئلتك"

def ask_question(user_question):
    if index is None:
        return "الرجاء رفع الكتب أولاً وتدريب النظام."
    rel_chunks = retrieve_relevant_chunks(user_question, index, chunks, None)
    answer = generate_answer(rel_chunks, user_question)
    return answer

# ====== Gradio Interface ======
upload = gr.File(file_types=[".pdf", ".doc", ".docx"], file_count="multiple")
train_btn = gr.Button("ابدأ التدريب")
train_output = gr.Textbox()
question_input = gr.Textbox(placeholder="اكتب سؤالك هنا باللغة العربية")
answer_output = gr.Textbox()
ask_btn = gr.Button("أرسل السؤال")

with gr.Blocks() as demo:
    gr.Markdown("# 🧠 محاكاة دماغ المؤلف - نظام ذكي للإجابة على الأسئلة من كتبك بالعربية")
    upload.render()
    train_btn.render()
    train_output.render()
    question_input.render()
    ask_btn.render()
    answer_output.render()

    train_btn.click(upload_and_train, inputs=[upload], outputs=[train_output])
    ask_btn.click(ask_question, inputs=[question_input], outputs=[answer_output])

# Launch
if __name__ == "__main__":
    demo.launch()