app.py

import os
import time
import streamlit as st
from langchain.chat_models import ChatOpenAI

from transformers import AutoTokenizer, AutoModel
from langchain.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.schema import Document as LangchainDocument
from langchain.chains import RetrievalQA
from langchain.llms import OpenAI
import torch
from langchain_core.retrievers import BaseRetriever
from langchain_core.documents import Document
from typing import List
from pydantic import Field
from sentence_transformers import SentenceTransformer
import numpy as np

# ----------------- تنظیمات صفحه -----------------
st.set_page_config(page_title="چت‌بات ارتش - فقط از PDF", page_icon="🪖", layout="wide")

# ----------------- بارگذاری مدل FarsiBERT -----------------
# model_name = "HooshvareLab/bert-fa-zwnj-base"
# tokenizer = AutoTokenizer.from_pretrained(model_name)
# model = AutoModel.from_pretrained(model_name)

# ----------------- لود PDF و ساخت ایندکس -----------------

@st.cache_resource
def build_pdf_index():
    with st.spinner('📄 در حال پردازش فایل PDF...'):
        # بارگذاری فایل
        loader = PyPDFLoader("test1.pdf")
        pages = loader.load()

        # تکه‌تکه کردن متن
        splitter = RecursiveCharacterTextSplitter(
            chunk_size=500,
            chunk_overlap=50
        )

        texts = []
        for page in pages:
            texts.extend(splitter.split_text(page.page_content))

        documents = [LangchainDocument(page_content=t) for t in texts]

        # مدل‌های Embedding
        sentence_model = SentenceTransformer('paraphrase-multilingual-MiniLM-L12-v2')

        embeddings = []

        # پروگرس بار اضافه می‌کنیم
        progress_bar = st.progress(0)
        total_docs = len(documents)

        # فقط از SentenceTransformer استفاده می‌کنیم
        for i, doc in enumerate(documents):
            batch_embedding = sentence_model.encode(doc.page_content, convert_to_numpy=True)
            embeddings.append(batch_embedding)

            # به‌روزرسانی پروگرس بار
            progress_bar.progress((i + 1) / total_docs)

        # اطمینان از اینکه خروجی NumpyArray است
        embeddings = np.array(embeddings)

        return documents, embeddings


# ----------------- تعریف LLM از Groq -----------------
# groq_api_key = "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"

# به جای OpenAI اینو بذار:
llm = ChatOpenAI(
    base_url="https://api.together.xyz/v1",
    api_key='0291f33aee03412a47fa5d8e562e515182dcc5d9aac5a7fb5eefdd1759005979',
    model="meta-llama/Llama-3.3-70B-Instruct-Turbo-Free"
)

# ----------------- تعریف SimpleRetriever -----------------
class SimpleRetriever(BaseRetriever):
    documents: List[Document] = Field(...)
    embeddings: List = Field(...)

    def _get_relevant_documents(self, query: str) -> List[Document]:
        # فقط از sentence_model استفاده می‌کنیم
        sentence_model = SentenceTransformer('paraphrase-multilingual-MiniLM-L12-v2')
        query_embedding = sentence_model.encode(query, convert_to_numpy=True)

        similarities = []
        for doc_embedding in self.embeddings:
            similarity = (query_embedding * doc_embedding).sum()
            similarities.append(similarity)

        ranked_docs = sorted(zip(similarities, self.documents), reverse=True)
        return [doc for _, doc in ranked_docs[:5]]

# ----------------- ساخت Index -----------------
documents, embeddings = build_pdf_index()
retriever = SimpleRetriever(documents=documents, embeddings=embeddings)

# ----------------- ساخت Chain -----------------
chain = RetrievalQA.from_chain_type(
    llm=llm,
    retriever=retriever,
    chain_type="stuff",
    input_key="question"
)

# ----------------- استیت برای چت -----------------
if 'messages' not in st.session_state:
    st.session_state.messages = []

if 'pending_prompt' not in st.session_state:
    st.session_state.pending_prompt = None

# ----------------- نمایش پیام‌های قبلی -----------------
for msg in st.session_state.messages:
    with st.chat_message(msg['role']):
        st.markdown(f"🗨️ {msg['content']}", unsafe_allow_html=True)

# ----------------- ورودی چت -----------------
prompt = st.chat_input("سوالی در مورد فایل بپرس...")

if prompt:
    st.session_state.messages.append({'role': 'user', 'content': prompt})
    st.session_state.pending_prompt = prompt
    st.rerun()

# ----------------- پاسخ مدل -----------------
if st.session_state.pending_prompt:
    with st.chat_message('ai'):
        thinking = st.empty()
        thinking.markdown("🤖 در حال فکر کردن از روی PDF...")

        try:
            response = chain.run(f"سوال: {st.session_state.pending_prompt}")
            answer = response.strip()
        except Exception as e:
            answer = f"خطا در پاسخ‌دهی: {str(e)}"

        thinking.empty()

        full_response = ""
        placeholder = st.empty()
        for word in answer.split():
            full_response += word + " "
            placeholder.markdown(full_response + "▌")
            time.sleep(0.03)

        placeholder.markdown(full_response)
        st.session_state.messages.append({'role': 'ai', 'content': full_response})
        st.session_state.pending_prompt = None