File size: 10,107 Bytes
f19a164 296403d f19a164 296403d f19a164 296403d f19a164 296403d f19a164 296403d f19a164 296403d f19a164 296403d f19a164 296403d f19a164 296403d f19a164 296403d f19a164 296403d |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 |
import gradio as gr
import os
from langchain.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.vectorstores import Chroma
from langchain.chains import ConversationalRetrievalChain
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.llms import HuggingFaceHub
from langchain.memory import ConversationBufferMemory
import chromadb
from transformers import AutoTokenizer
import transformers
import torch
# Constants and configuration
list_llm = [
"mistralai/Mixtral-8x7B-Instruct-v0.1", "mistralai/Mistral-7B-Instruct-v0.2",
"mistralai/Mistral-7B-Instruct-v0.1", "HuggingFaceH4/zephyr-7b-beta",
"meta-llama/Llama-2-7b-chat-hf", "microsoft/phi-2",
"TinyLlama/TinyLlama-1.1B-Chat-v1.0", "mosaicml/mpt-7b-instruct",
"tiiuae/falcon-7b-instruct", "google/flan-t5-xxl"
]
list_llm_simple = [os.path.basename(llm) for llm in list_llm]
# Function placeholders (actual function implementations from the original script)
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
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
def load_db():
embedding = HuggingFaceEmbeddings()
vectordb = Chroma(embedding_function=embedding)
return vectordb
def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db):
if llm_model == "mistralai/Mixtral-8x7B-Instruct-v0.1":
llm = HuggingFaceHub(
repo_id=llm_model,
model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k, "load_in_8bit": True}
)
else:
llm = HuggingFaceHub(
repo_id=llm_model,
model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k}
)
memory = ConversationBufferMemory(
memory_key="chat_history",
output_key='answer',
return_messages=True
)
retriever = vector_db.as_retriever()
qa_chain = ConversationalRetrievalChain.from_llm(
llm,
retriever=retriever,
chain_type="stuff",
memory=memory,
return_source_documents=True,
return_generated_question=False,
)
return qa_chain
def initialize_database(list_file_obj, chunk_size, chunk_overlap):
list_file_path = [x.name for x in list_file_obj if x is not None]
collection_name = os.path.basename(list_file_path[0]).replace(" ","-")[:50]
doc_splits = load_doc(list_file_path, chunk_size, chunk_overlap)
vector_db = create_db(doc_splits, collection_name)
return vector_db, collection_name, "Complete!"
def initialize_LLM(llm_option, llm_temperature, max_tokens, top_k, vector_db):
llm_name = list_llm[llm_option]
qa_chain = initialize_llmchain(llm_name, llm_temperature, max_tokens, top_k, vector_db)
return qa_chain, "Complete!"
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
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"]
response_sources = response["source_documents"]
response_source1 = response_sources[0].page_content.strip()
response_source2 = response_sources[1].page_content.strip()
response_source1_page = response_sources[0].metadata["page"] + 1
response_source2_page = response_sources[1].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
def upload_file(file_obj):
list_file_path = []
for idx, file in enumerate(file_obj):
file_path = file_obj.name
list_file_path.append(file_path)
return list_file_path
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 (powered by LangChain and open-source LLMs)</center></h2>
<h3>Ask any questions about your PDF documents, along with follow-ups</h3>
<b>Note:</b> This AI assistant performs retrieval-augmented generation from your PDF documents. \
When generating answers, it takes past questions into account (via conversational memory), and includes document references for clarity purposes.</i>
<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 an output.<br>
""")
with gr.Tab("Step 1 - Document pre-processing"):
with gr.Row():
document = gr.Files(height=100, file_count="multiple", file_types=["pdf"], interactive=True, label="Upload your PDF documents (single or multiple)")
# upload_btn = gr.UploadButton("Loading document...", height=100, file_count="multiple", file_types=["pdf"], scale=1)
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 2 - QA chain initialization"):
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 = 0.0, 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 3 - Conversation with 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():
msg = gr.Textbox(placeholder="Type message", container=True)
with gr.Row():
submit_btn = gr.Button("Submit")
clear_btn = gr.ClearButton([msg, chatbot])
# Preprocessing events
#upload_btn.upload(upload_file, inputs=[upload_btn], outputs=[document])
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], \
inputs=None, \
outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page], \
queue=False)
# Chatbot events
msg.submit(conversation, \
inputs=[qa_chain, msg, chatbot], \
outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_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], \
queue=False)
clear_btn.click(lambda:[None,"",0,"",0], \
inputs=None, \
outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page], \
queue=False)
demo.queue().launch(debug=True)
if __name__ == "__main__":
demo() |