file_chat_bot / app.py
taewon99's picture
Update app.py
67cdffd
raw history blame
No virus
8.32 kB
import streamlit as st
from langchain.text_splitter import (
CharacterTextSplitter,
RecursiveCharacterTextSplitter,
)
from langchain.vectorstores import FAISS
from langchain.embeddings import (
HuggingFaceEmbeddings,
) # General embeddings from HuggingFace models.
from langchain.memory import ConversationBufferMemory
from langchain.chains import ConversationalRetrievalChain
from htmlTemplates import css, bot_template, user_template
from langchain.llms import LlamaCpp # For loading transformer models.
from langchain.document_loaders import PyPDFLoader, TextLoader, JSONLoader, CSVLoader
import tempfile # μž„μ‹œ νŒŒμΌμ„ μƒμ„±ν•˜κΈ° μœ„ν•œ λΌμ΄λΈŒλŸ¬λ¦¬μž…λ‹ˆλ‹€.
import os
from huggingface_hub import hf_hub_download # Hugging Face Hubμ—μ„œ λͺ¨λΈμ„ λ‹€μš΄λ‘œλ“œν•˜κΈ° μœ„ν•œ ν•¨μˆ˜μž…λ‹ˆλ‹€.
# PDF λ¬Έμ„œλ‘œλΆ€ν„° ν…μŠ€νŠΈλ₯Ό μΆ”μΆœν•˜λŠ” ν•¨μˆ˜μž…λ‹ˆλ‹€.
def get_pdf_text(pdf_docs):
temp_dir = tempfile.TemporaryDirectory() # μž„μ‹œ 디렉토리λ₯Ό μƒμ„±ν•©λ‹ˆλ‹€.
temp_filepath = os.path.join(temp_dir.name, pdf_docs.name) # μž„μ‹œ 파일 경둜λ₯Ό μƒμ„±ν•©λ‹ˆλ‹€.
with open(temp_filepath, "wb") as f: # μž„μ‹œ νŒŒμΌμ„ λ°”μ΄λ„ˆλ¦¬ μ“°κΈ° λͺ¨λ“œλ‘œ μ—½λ‹ˆλ‹€.
f.write(pdf_docs.getvalue()) # PDF λ¬Έμ„œμ˜ λ‚΄μš©μ„ μž„μ‹œ νŒŒμΌμ— μ”λ‹ˆλ‹€.
pdf_loader = PyPDFLoader(temp_filepath) # PyPDFLoaderλ₯Ό μ‚¬μš©ν•΄ PDFλ₯Ό λ‘œλ“œν•©λ‹ˆλ‹€.
pdf_doc = pdf_loader.load() # ν…μŠ€νŠΈλ₯Ό μΆ”μΆœν•©λ‹ˆλ‹€.
return pdf_doc # μΆ”μΆœν•œ ν…μŠ€νŠΈλ₯Ό λ°˜ν™˜ν•©λ‹ˆλ‹€.
# 과제
# μ•„λž˜ ν…μŠ€νŠΈ μΆ”μΆœ ν•¨μˆ˜λ₯Ό μž‘μ„±
def get_text_file(docs):
temp_dir = tempfile.TemporaryDirectory() # μž„μ‹œ 디렉토리λ₯Ό μƒμ„±ν•©λ‹ˆλ‹€.
temp_filepath = os.path.join(temp_dir.name, docs.name) # μž„μ‹œ 파일 경둜λ₯Ό μƒμ„±ν•©λ‹ˆλ‹€.
with open(temp_filepath, "wb") as f: # μž„μ‹œ νŒŒμΌμ„ λ°”μ΄λ„ˆλ¦¬ μ“°κΈ° λͺ¨λ“œλ‘œ μ—½λ‹ˆλ‹€.
f.write(docs.getvalue()) # ν…μŠ€νŠΈ 파일의 λ‚΄μš©μ„ μž„μ‹œ νŒŒμΌμ— μ”λ‹ˆλ‹€.
text_loader = TextLoader(temp_filepath) # TextLoaderλ₯Ό μ‚¬μš©ν•΄ ν…μŠ€νŠΈλ₯Ό λ‘œλ“œν•©λ‹ˆλ‹€.
text_doc = text_loader.load() # ν…μŠ€νŠΈλ₯Ό μΆ”μΆœν•©λ‹ˆλ‹€.
return text_doc # μΆ”μΆœν•œ ν…μŠ€νŠΈλ₯Ό λ°˜ν™˜ν•©λ‹ˆλ‹€.
def get_csv_file(docs):
temp_dir = tempfile.TemporaryDirectory() # μž„μ‹œ 디렉토리λ₯Ό μƒμ„±ν•©λ‹ˆλ‹€.
temp_filepath = os.path.join(temp_dir.name, docs.name) # μž„μ‹œ 파일 경둜λ₯Ό μƒμ„±ν•©λ‹ˆλ‹€.
with open(temp_filepath, "wb") as f: # μž„μ‹œ νŒŒμΌμ„ λ°”μ΄λ„ˆλ¦¬ μ“°κΈ° λͺ¨λ“œλ‘œ μ—½λ‹ˆλ‹€.
f.write(docs.getvalue()) # CSV 파일의 λ‚΄μš©μ„ μž„μ‹œ νŒŒμΌμ— μ”λ‹ˆλ‹€.
csv_loader = CSVLoader(temp_filepath) # CSVLoaderλ₯Ό μ‚¬μš©ν•΄ CSVλ₯Ό λ‘œλ“œν•©λ‹ˆλ‹€.
csv_doc = csv_loader.load() # ν…μŠ€νŠΈλ₯Ό μΆ”μΆœν•©λ‹ˆλ‹€.
return csv_doc # μΆ”μΆœν•œ ν…μŠ€νŠΈλ₯Ό λ°˜ν™˜ν•©λ‹ˆλ‹€.
def get_json_file(docs):
temp_dir = tempfile.TemporaryDirectory() # μž„μ‹œ 디렉토리λ₯Ό μƒμ„±ν•©λ‹ˆλ‹€.
temp_filepath = os.path.join(temp_dir.name, docs.name) # μž„μ‹œ 파일 경둜λ₯Ό μƒμ„±ν•©λ‹ˆλ‹€.
with open(temp_filepath, "wb") as f: # μž„μ‹œ νŒŒμΌμ„ λ°”μ΄λ„ˆλ¦¬ μ“°κΈ° λͺ¨λ“œλ‘œ μ—½λ‹ˆλ‹€.
f.write(docs.getvalue()) # JSON 파일의 λ‚΄μš©μ„ μž„μ‹œ νŒŒμΌμ— μ”λ‹ˆλ‹€.
json_loader = JSONLoader(temp_filepath) # JSONLoaderλ₯Ό μ‚¬μš©ν•΄ JSON을 λ‘œλ“œν•©λ‹ˆλ‹€.
json_doc = json_loader.load() # ν…μŠ€νŠΈλ₯Ό μΆ”μΆœν•©λ‹ˆλ‹€.
return json_doc # μΆ”μΆœν•œ ν…μŠ€νŠΈλ₯Ό λ°˜ν™˜ν•©λ‹ˆλ‹€.
# λ¬Έμ„œλ“€μ„ μ²˜λ¦¬ν•˜μ—¬ ν…μŠ€νŠΈ 청크둜 λ‚˜λˆ„λŠ” ν•¨μˆ˜μž…λ‹ˆλ‹€.
def get_text_chunks(documents):
text_splitter = RecursiveCharacterTextSplitter(
chunk_size=1000, # 청크의 크기λ₯Ό μ§€μ •ν•©λ‹ˆλ‹€.
chunk_overlap=200, # 청크 μ‚¬μ΄μ˜ 쀑볡을 μ§€μ •ν•©λ‹ˆλ‹€.
length_function=len, # ν…μŠ€νŠΈμ˜ 길이λ₯Ό μΈ‘μ •ν•˜λŠ” ν•¨μˆ˜λ₯Ό μ§€μ •ν•©λ‹ˆλ‹€.
)
documents = text_splitter.split_documents(documents) # λ¬Έμ„œλ“€μ„ 청크둜 λ‚˜λˆ•λ‹ˆλ‹€.
return documents # λ‚˜λˆˆ 청크λ₯Ό λ°˜ν™˜ν•©λ‹ˆλ‹€.
# ν…μŠ€νŠΈ μ²­ν¬λ“€λ‘œλΆ€ν„° 벑터 μŠ€ν† μ–΄λ₯Ό μƒμ„±ν•˜λŠ” ν•¨μˆ˜μž…λ‹ˆλ‹€.
def get_vectorstore(text_chunks):
# μ›ν•˜λŠ” μž„λ² λ”© λͺ¨λΈμ„ λ‘œλ“œν•©λ‹ˆλ‹€.
embeddings = HuggingFaceEmbeddings(
model_name="bert-base-multilingual-cased",
model_kwargs={"device": "cpu"},
) # μž„λ² λ”© λͺ¨λΈμ„ μ„€μ •ν•©λ‹ˆλ‹€.
vectorstore = FAISS.from_documents(text_chunks, embeddings) # FAISS 벑터 μŠ€ν† μ–΄λ₯Ό μƒμ„±ν•©λ‹ˆλ‹€.
return vectorstore # μƒμ„±λœ 벑터 μŠ€ν† μ–΄λ₯Ό λ°˜ν™˜ν•©λ‹ˆλ‹€.
def get_conversation_chain(vectorstore):
model_name_or_path = "TheBloke/Llama-2-7B-chat-GGUF"
model_basename = "llama-2-7b-chat.Q2_K.gguf"
model_path = hf_hub_download(repo_id=model_name_or_path, filename=model_basename)
llm = LlamaCpp(
model_path=model_path,
n_ctx=4086,
input={"temperature": 0.75, "max_length": 2000, "top_p": 1},
verbose=True,
)
# λŒ€ν™” 기둝을 μ €μž₯ν•˜κΈ° μœ„ν•œ λ©”λͺ¨λ¦¬λ₯Ό μƒμ„±ν•©λ‹ˆλ‹€.
memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True)
# λŒ€ν™” 검색 체인을 μƒμ„±ν•©λ‹ˆλ‹€.
conversation_chain = ConversationalRetrievalChain.from_llm(
llm=llm, retriever=vectorstore.as_retriever(), memory=memory
)
return conversation_chain # μƒμ„±λœ λŒ€ν™” 체인을 λ°˜ν™˜ν•©λ‹ˆλ‹€.
# μ‚¬μš©μž μž…λ ₯을 μ²˜λ¦¬ν•˜λŠ” ν•¨μˆ˜μž…λ‹ˆλ‹€.
def handle_userinput(user_question):
print("user_question => ", user_question)
# λŒ€ν™” 체인을 μ‚¬μš©ν•˜μ—¬ μ‚¬μš©μž μ§ˆλ¬Έμ— λŒ€ν•œ 응닡을 μƒμ„±ν•©λ‹ˆλ‹€.
response = st.session_state.conversation({"question": user_question})
# λŒ€ν™” 기둝을 μ €μž₯ν•©λ‹ˆλ‹€.
st.session_state.chat_history = response["chat_history"]
for i, message in enumerate(st.session_state.chat_history):
if i % 2 == 0:
st.write(
user_template.replace("{{MSG}}", message.content),
unsafe_allow_html=True,
)
else:
st.write(
bot_template.replace("{{MSG}}", message.content), unsafe_allow_html=True
)
def main():
# Streamlit UIλ₯Ό μ„€μ •ν•©λ‹ˆλ‹€.
st.title("λ¬Έμ„œμ— λŒ€ν•΄ μ§ˆλ¬Έν•˜λŠ” μ±„νŒ…")
st.sidebar.title("λ¬Έμ„œ μ—…λ‘œλ“œ")
st.sidebar.text("여기에 μ—¬λŸ¬λΆ„μ˜ λ¬Έμ„œλ₯Ό μ—…λ‘œλ“œν•˜μ„Έμš”.")
st.sidebar.text("PDF, TXT, CSV, JSON 파일 ν˜•μ‹μ„ μ§€μ›ν•©λ‹ˆλ‹€.")
st.write(css, unsafe_allow_html=True)
if "conversation" not in st.session_state:
st.session_state.conversation = None
if "chat_history" not in st.session_state:
st.session_state.chat_history = None
st.header("λ¬Έμ„œμ— λŒ€ν•΄ μ§ˆλ¬Έν•˜λŠ” μ±„νŒ…")
user_question = st.text_input("문선에 λŒ€ν•΄μ„œ μ§ˆλ¬Έμ„ ν•΄λ³΄μ„Έμš”")
if user_question:
handle_userinput(user_question)
with st.sidebar:
st.subheader("Your documents")
docs = st.file_uploader("여기에 λ¬Έμ„œλ₯Ό μ—…λ‘œλ“œν•˜κ³  '처리'λ₯Ό ν΄λ¦­ν•˜μ„Έμš”", accept_multiple_files=True)
if st.button("처리"):
with st.spinner("Processing"):
# get pdf text
doc_list = []
for file in docs:
print("file - type : ", file.type)
if file.type == "text/plain":
# file is .txt
doc_list.extend(get_text_file(file))
elif file.type in ["application/octet-stream", "application/pdf"]:
# file is .pdf
doc_list.extend(get_pdf_text(file))
elif file.type == "text/csv":
# file is .csv
doc_list.extend(get_csv_file(file))
elif file.type == "application/json":
# file is .json
doc_list.extend(get_json_file(file))
# get the text chunks
text_chunks = get_text_chunks(doc_list)
# create vector store
vectorstore = get_vectorstore(text_chunks)
# create conversation chain
st.session_state.conversation = get_conversation_chain(vectorstore)
if __name__ == "__main__":
main()