pdfchatbot / app.py
DHEIVER's picture
Update app.py
7e34d60 verified
raw
history blame
11.8 kB
import gradio as gr
import os
from langchain_community.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.vectorstores import Chroma
from langchain.chains import ConversationalRetrievalChain
from langchain_community.embeddings import HuggingFaceEmbeddings
from langchain_community.llms import HuggingFacePipeline
from langchain.chains import ConversationChain
from langchain.memory import ConversationBufferMemory
from langchain_community.llms import HuggingFaceEndpoint
from pathlib import Path
import chromadb
from unidecode import unidecode
from transformers import AutoTokenizer, pipeline
import transformers
import torch
import tqdm
import accelerate
import re
# Lista de modelos gratuitos que não exigem chave de API
list_llm = [
"mistralai/Mistral-7B-Instruct-v0.2",
"mistralai/Mistral-7B-Instruct-v0.1",
"google/flan-t5-xxl",
"HuggingFaceH4/zephyr-7b-beta",
"TinyLlama/TinyLlama-1.1B-Chat-v1.0",
"microsoft/phi-2"
]
list_llm_simple = [os.path.basename(llm) for llm in list_llm]
# Função para carregar o documento PDF e dividir em partes
def load_doc(list_file_path, chunk_size, chunk_overlap):
loaders = [PyPDFLoader(x) for x in list_file_path]
pages = []
for loader in loaders:
pages.extend(loader.load())
text_splitter = RecursiveCharacterTextSplitter(
chunk_size=chunk_size,
chunk_overlap=chunk_overlap
)
doc_splits = text_splitter.split_documents(pages)
return doc_splits
# Função para criar o banco de dados vetorial
def create_db(splits, collection_name):
embedding = HuggingFaceEmbeddings()
new_client = chromadb.EphemeralClient()
vectordb = Chroma.from_documents(
documents=splits,
embedding=embedding,
client=new_client,
collection_name=collection_name,
)
return vectordb
# Função para inicializar a cadeia de LLM
def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
progress(0.1, desc="Initializing HF tokenizer...")
# Carregar o tokenizer e o pipeline do modelo
tokenizer = AutoTokenizer.from_pretrained(llm_model)
progress(0.5, desc="Initializing HF pipeline...")
pipeline_model = transformers.pipeline(
"text-generation",
model=llm_model,
tokenizer=tokenizer,
torch_dtype=torch.bfloat16,
device_map="auto",
max_new_tokens=max_tokens,
do_sample=True,
top_k=top_k,
temperature=temperature,
)
llm = HuggingFacePipeline(pipeline=pipeline_model)
progress(0.75, desc="Defining buffer memory...")
memory = ConversationBufferMemory(
memory_key="chat_history",
output_key='answer',
return_messages=True
)
retriever = vector_db.as_retriever()
progress(0.8, desc="Defining retrieval chain...")
qa_chain = ConversationalRetrievalChain.from_llm(
llm,
retriever=retriever,
chain_type="stuff",
memory=memory,
return_source_documents=True,
verbose=False,
)
progress(0.9, desc="Done!")
return qa_chain
# Função para gerar o nome da coleção do banco de dados vetorial
def create_collection_name(filepath):
collection_name = Path(filepath).stem
collection_name = collection_name.replace(" ", "-")
collection_name = unidecode(collection_name)
collection_name = re.sub('[^A-Za-z0-9]+', '-', collection_name)
collection_name = collection_name[:50]
if len(collection_name) < 3:
collection_name = collection_name + 'xyz'
if not collection_name[0].isalnum():
collection_name = 'A' + collection_name[1:]
if not collection_name[-1].isalnum():
collection_name = collection_name[:-1] + 'Z'
return collection_name
# Função para inicializar o banco de dados
def initialize_database(list_file_obj, chunk_size, chunk_overlap, progress=gr.Progress()):
list_file_path = [x.name for x in list_file_obj if x is not None]
progress(0.1, desc="Creating collection name...")
collection_name = create_collection_name(list_file_path[0])
progress(0.25, desc="Loading document...")
doc_splits = load_doc(list_file_path, chunk_size, chunk_overlap)
progress(0.5, desc="Generating vector database...")
vector_db = create_db(doc_splits, collection_name)
progress(0.9, desc="Done!")
return vector_db, collection_name, "Complete!"
# Função para inicializar a cadeia de QA
def initialize_LLM(llm_option, llm_temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
llm_name = list_llm[llm_option]
print("llm_name: ", llm_name)
qa_chain = initialize_llmchain(llm_name, llm_temperature, max_tokens, top_k, vector_db, progress)
return qa_chain, "Complete!"
# Função para formatar o histórico de conversa
def format_chat_history(message, chat_history):
formatted_chat_history = []
for user_message, bot_message in chat_history:
formatted_chat_history.append(f"User: {user_message}")
formatted_chat_history.append(f"Assistant: {bot_message}")
return formatted_chat_history
# Função para gerar a conversa
def conversation(qa_chain, message, history):
formatted_chat_history = format_chat_history(message, history)
response = qa_chain({"question": message, "chat_history": formatted_chat_history})
response_answer = response["answer"]
if response_answer.find("Helpful Answer:") != -1:
response_answer = response_answer.split("Helpful Answer:")[-1]
response_sources = response["source_documents"]
response_source1 = response_sources[0].page_content.strip()
response_source2 = response_sources[1].page_content.strip()
response_source3 = response_sources[2].page_content.strip()
response_source1_page = response_sources[0].metadata["page"] + 1
response_source2_page = response_sources[1].metadata["page"] + 1
response_source3_page = response_sources[2].metadata["page"] + 1
new_history = history + [(message, response_answer)]
return qa_chain, gr.update(value=""), new_history, response_source1, response_source1_page, response_source2, response_source2_page, response_source3, response_source3_page
# Função principal para rodar a interface
def demo():
with gr.Blocks(theme="base") as demo:
vector_db = gr.State()
qa_chain = gr.State()
collection_name = gr.State()
gr.Markdown(
"""<center><h2>PDF-based chatbot</center></h2>
<h3>Ask any questions about your PDF documents</h3>""")
gr.Markdown(
"""<b>Note:</b> This AI assistant, using Langchain and open-source LLMs, performs retrieval-augmented generation (RAG) from your PDF documents. \
The user interface explicitely shows multiple steps to help understand the RAG workflow.
This chatbot takes past questions into account when generating answers (via conversational memory), and includes document references for clarity purposes.<br>
<br><b>Warning:</b> This space uses the free CPU Basic hardware from Hugging Face. Some steps and LLM models used below (free inference endpoints) can take some time to generate a reply.
""")
with gr.Tab("Step 1 - Upload PDF"):
with gr.Row():
document = gr.Files(height=100, file_count="multiple", file_types=["pdf"], interactive=True, label="Upload your PDF documents (single or multiple)")
with gr.Tab("Step 2 - Process document"):
with gr.Row():
db_btn = gr.Radio(["ChromaDB"], label="Vector database type", value="ChromaDB", type="index", info="Choose your vector database")
with gr.Accordion("Advanced options - Document text splitter", open=False):
with gr.Row():
slider_chunk_size = gr.Slider(minimum=100, maximum=1000, value=600, step=20, label="Chunk size", info="Chunk size", interactive=True)
with gr.Row():
slider_chunk_overlap = gr.Slider(minimum=10, maximum=200, value=40, step=10, label="Chunk overlap", info="Chunk overlap", interactive=True)
with gr.Row():
db_progress = gr.Textbox(label="Vector database initialization", value="None")
with gr.Row():
db_btn = gr.Button("Generate vector database")
with gr.Tab("Step 3 - Initialize QA chain"):
with gr.Row():
llm_btn = gr.Radio(list_llm_simple, label="LLM models", value=list_llm_simple[0], type="index", info="Choose your LLM model")
with gr.Accordion("Advanced options - LLM model", open=False):
with gr.Row():
slider_temperature = gr.Slider(minimum=0.01, maximum=1.0, value=0.7, step=0.1, label="Temperature", info="Model temperature", interactive=True)
with gr.Row():
slider_maxtokens = gr.Slider(minimum=224, maximum=4096, value=1024, step=32, label="Max Tokens", info="Model max tokens", interactive=True)
with gr.Row():
slider_topk = gr.Slider(minimum=1, maximum=10, value=3, step=1, label="top-k samples", info="Model top-k samples", interactive=True)
with gr.Row():
llm_progress = gr.Textbox(value="None", label="QA chain initialization")
with gr.Row():
qachain_btn = gr.Button("Initialize Question Answering chain")
with gr.Tab("Step 4 - Chatbot"):
chatbot = gr.Chatbot(height=300)
with gr.Accordion("Advanced - Document references", open=False):
with gr.Row():
doc_source1 = gr.Textbox(label="Reference 1", lines=2, container=True, scale=20)
source1_page = gr.Number(label="Page", scale=1)
with gr.Row():
doc_source2 = gr.Textbox(label="Reference 2", lines=2, container=True, scale=20)
source2_page = gr.Number(label="Page", scale=1)
with gr.Row():
doc_source3 = gr.Textbox(label="Reference 3", lines=2, container=True, scale=20)
source3_page = gr.Number(label="Page", scale=1)
with gr.Row():
msg = gr.Textbox(placeholder="Type message (e.g. 'What is this document about?')", container=True)
with gr.Row():
submit_btn = gr.Button("Submit message")
clear_btn = gr.ClearButton([msg, chatbot], value="Clear conversation")
# Eventos de pré-processamento
db_btn.click(initialize_database, inputs=[document, slider_chunk_size, slider_chunk_overlap], outputs=[vector_db, collection_name, db_progress])
qachain_btn.click(initialize_LLM, inputs=[llm_btn, slider_temperature, slider_maxtokens, slider_topk, vector_db], outputs=[qa_chain, llm_progress]).then(lambda:[None,"",0,"",0,"",0], inputs=None, outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], queue=False)
# Eventos do chatbot
msg.submit(conversation, inputs=[qa_chain, msg, chatbot], outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], queue=False)
submit_btn.click(conversation, inputs=[qa_chain, msg, chatbot], outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], queue=False)
clear_btn.click(lambda:[None,"",0,"",0,"",0], inputs=None, outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], queue=False)
demo.queue().launch(debug=True)
if __name__ == "__main__":
demo()