랭체인 PDF 요약기 만들기(eeve:q4)

[주인장]

import os

from PyPDF2 import PdfReader

import streamlit as st

from langchain.text_splitter import CharacterTextSplitter

from langchain.vectorstores import FAISS

from langchain.chains.question_answering import load_qa_chain

from langchain_community.embeddings import HuggingFaceEmbeddings

from langchain_community.chat_models import ChatOllama

def process_text(text):

    # 텍스트를 청크로 분할

    text_splitter = CharacterTextSplitter(

        separator="\n",

        chunk_size=1000,

        chunk_overlap=200,

        length_function=len

    )

    chunks = text_splitter.split_text(text)

    # HuggingFace 임베딩 사용 (Ollama에서는 직접 임베딩 미지원)

    embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2")  # 한국어용 모델

    documents = FAISS.from_texts(chunks, embeddings)

    return documents

def main():

  st.title("📄PDF 요약하기 (eeve:q4)")

  st.divider()

  pdf = st.file_uploader('PDF 파일을 업로드해주세요', type='pdf')

  if pdf is not None:

    pdf_reader = PdfReader(pdf)

    text = ""

    for page in pdf_reader.pages:

      text += page.extract_text()

    documents = process_text(text)

    query = "업로드된 PDF 파일의 내용을 약 3~5문장으로 요약해주세요."

    if query:

      docs = documents.similarity_search(query)

      # Ollama를 사용하는 LLM (eeve:q4 모델)

      llm = ChatOllama(model="eeve:q4", temperature=0.1)

      chain = load_qa_chain(llm, chain_type='stuff')

      response = chain.run(input_documents=docs, question=query)

      st.subheader('--요약 결과--:')

      st.write(response)

if __name__ == '__main__':

  main()