clementsan
commited on
Commit
•
5be8df6
1
Parent(s):
529bde4
Add PDF chatbot application
Browse files
app.py
CHANGED
@@ -1,7 +1,239 @@
|
|
1 |
import gradio as gr
|
|
|
2 |
|
3 |
-
|
4 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
5 |
|
6 |
-
|
7 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
2 |
+
import os
|
3 |
|
4 |
+
from langchain.document_loaders import PyPDFLoader
|
5 |
+
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
6 |
+
from langchain.vectorstores import Chroma
|
7 |
+
from langchain.chains import ConversationalRetrievalChain
|
8 |
+
from langchain.embeddings import HuggingFaceEmbeddings
|
9 |
+
from langchain.llms import HuggingFacePipeline
|
10 |
+
from langchain.chains import ConversationChain
|
11 |
+
from langchain.memory import ConversationBufferMemory
|
12 |
+
from langchain.llms import HuggingFaceHub
|
13 |
|
14 |
+
from transformers import AutoTokenizer
|
15 |
+
import transformers
|
16 |
+
import torch
|
17 |
+
import tqdm
|
18 |
+
import accelerate
|
19 |
+
|
20 |
+
|
21 |
+
default_persist_directory = './chroma_HF/'
|
22 |
+
default_llm_name1 = "tiiuae/falcon-7b-instruct"
|
23 |
+
default_llm_name2 = "google/flan-t5-xxl"
|
24 |
+
default_llm_name3 = "mosaicml/mpt-7b-instruct"
|
25 |
+
default_llm_name4 = "meta-llama/Llama-2-7b-chat-hf"
|
26 |
+
default_llm_name5 = "mistralai/Mistral-7B-Instruct-v0.1"
|
27 |
+
list_llm = [default_llm_name1, default_llm_name2, default_llm_name3, default_llm_name4, default_llm_name5]
|
28 |
+
|
29 |
+
|
30 |
+
# Load PDF document and create doc splits
|
31 |
+
def load_doc(list_file_path, chunk_size, chunk_overlap):
|
32 |
+
# Processing for one document only
|
33 |
+
# loader = PyPDFLoader(file_path)
|
34 |
+
# pages = loader.load()
|
35 |
+
loaders = [PyPDFLoader(x) for x in list_file_path]
|
36 |
+
pages = []
|
37 |
+
for loader in loaders:
|
38 |
+
pages.extend(loader.load())
|
39 |
+
# text_splitter = RecursiveCharacterTextSplitter(chunk_size = 600, chunk_overlap = 50)
|
40 |
+
text_splitter = RecursiveCharacterTextSplitter(
|
41 |
+
chunk_size = chunk_size,
|
42 |
+
chunk_overlap = chunk_overlap)
|
43 |
+
doc_splits = text_splitter.split_documents(pages)
|
44 |
+
return doc_splits
|
45 |
+
|
46 |
+
|
47 |
+
# Create vector database
|
48 |
+
def create_db(splits):
|
49 |
+
embedding = HuggingFaceEmbeddings()
|
50 |
+
vectordb = Chroma.from_documents(
|
51 |
+
documents=splits,
|
52 |
+
embedding=embedding,
|
53 |
+
persist_directory=default_persist_directory
|
54 |
+
)
|
55 |
+
return vectordb
|
56 |
+
|
57 |
+
|
58 |
+
# Load vector database
|
59 |
+
def load_db():
|
60 |
+
embedding = HuggingFaceEmbeddings()
|
61 |
+
vectordb = Chroma(
|
62 |
+
persist_directory=default_persist_directory,
|
63 |
+
embedding_function=embedding)
|
64 |
+
return vectordb
|
65 |
+
|
66 |
+
|
67 |
+
# Initialize langchain LLM chain
|
68 |
+
def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
|
69 |
+
progress(0.1, desc="Initializing HF tokenizer...")
|
70 |
+
# HuggingFacePipeline uses local model
|
71 |
+
# Warning: it will download model locally...
|
72 |
+
# tokenizer=AutoTokenizer.from_pretrained(llm_model)
|
73 |
+
# progress(0.5, desc="Initializing HF pipeline...")
|
74 |
+
# pipeline=transformers.pipeline(
|
75 |
+
# "text-generation",
|
76 |
+
# model=llm_model,
|
77 |
+
# tokenizer=tokenizer,
|
78 |
+
# torch_dtype=torch.bfloat16,
|
79 |
+
# trust_remote_code=True,
|
80 |
+
# device_map="auto",
|
81 |
+
# # max_length=1024,
|
82 |
+
# max_new_tokens=max_tokens,
|
83 |
+
# do_sample=True,
|
84 |
+
# top_k=top_k,
|
85 |
+
# num_return_sequences=1,
|
86 |
+
# eos_token_id=tokenizer.eos_token_id
|
87 |
+
# )
|
88 |
+
# llm = HuggingFacePipeline(pipeline=pipeline, model_kwargs={'temperature': temperature})
|
89 |
+
|
90 |
+
# HuggingFaceHub uses HF inference endpoints
|
91 |
+
progress(0.5, desc="Initializing HF Hub...")
|
92 |
+
llm = HuggingFaceHub(
|
93 |
+
repo_id=llm_model,
|
94 |
+
model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k}
|
95 |
+
)
|
96 |
+
|
97 |
+
progress(0.5, desc="Defining buffer memory...")
|
98 |
+
memory = ConversationBufferMemory(
|
99 |
+
memory_key="chat_history",
|
100 |
+
return_messages=True
|
101 |
+
)
|
102 |
+
# retriever=vector_db.as_retriever(search_type="similarity", search_kwargs={'k': 3})
|
103 |
+
retriever=vector_db.as_retriever()
|
104 |
+
progress(0.8, desc="Defining retrieval chain...")
|
105 |
+
global qa_chain
|
106 |
+
qa_chain = ConversationalRetrievalChain.from_llm(
|
107 |
+
llm,
|
108 |
+
retriever=retriever,
|
109 |
+
chain_type="stuff",
|
110 |
+
memory=memory,
|
111 |
+
# combine_docs_chain_kwargs={"prompt": your_prompt})
|
112 |
+
# return_source_documents=True,
|
113 |
+
# return_generated_question=True,
|
114 |
+
# verbose=True,
|
115 |
+
)
|
116 |
+
progress(0.9, desc="Done!")
|
117 |
+
# return qa_chain
|
118 |
+
|
119 |
+
|
120 |
+
# Initialize all elements
|
121 |
+
def initialize_database(list_file_obj, chunk_size, chunk_overlap, progress=gr.Progress()):
|
122 |
+
# Create list of documents (when valid)
|
123 |
+
#file_path = file_obj.name
|
124 |
+
list_file_path = [x.name for x in list_file_obj if x is not None]
|
125 |
+
print('list_file_path', list_file_path)
|
126 |
+
progress(0.25, desc="Loading document...")
|
127 |
+
# Load document and create splits
|
128 |
+
doc_splits = load_doc(list_file_path, chunk_size, chunk_overlap)
|
129 |
+
# Create or load Vector database
|
130 |
+
progress(0.5, desc="Generating vector database...")
|
131 |
+
# global vector_db
|
132 |
+
vector_db = create_db(doc_splits)
|
133 |
+
progress(0.9, desc="Done!")
|
134 |
+
return vector_db, "Complete!"
|
135 |
+
#return qa_chain
|
136 |
+
|
137 |
+
|
138 |
+
def initialize_LLM(llm_option, llm_temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
|
139 |
+
print("llm_option",llm_option)
|
140 |
+
llm_name = list_llm[llm_option]
|
141 |
+
print("llm_name",llm_name)
|
142 |
+
initialize_llmchain(llm_name, llm_temperature, max_tokens, top_k, vector_db, progress)
|
143 |
+
return "Complete!"
|
144 |
+
#return qa_chain
|
145 |
+
|
146 |
+
|
147 |
+
def format_chat_history(message, chat_history):
|
148 |
+
formatted_chat_history = []
|
149 |
+
for user_message, bot_message in chat_history:
|
150 |
+
formatted_chat_history.append(f"User: {user_message}")
|
151 |
+
formatted_chat_history.append(f"Assistant: {bot_message}")
|
152 |
+
return formatted_chat_history
|
153 |
+
|
154 |
+
|
155 |
+
def conversation(message, history):
|
156 |
+
formatted_chat_history = format_chat_history(message, history)
|
157 |
+
#print("formatted_chat_history",formatted_chat_history)
|
158 |
+
|
159 |
+
# Generate response using QA chain
|
160 |
+
response = qa_chain({"question": message, "chat_history": formatted_chat_history})
|
161 |
+
# return response['answer']
|
162 |
+
|
163 |
+
# Append user message and response to chat history
|
164 |
+
new_history = history + [(message, response["answer"])]
|
165 |
+
return gr.update(value=""), new_history
|
166 |
+
|
167 |
+
|
168 |
+
def upload_file(file_obj):
|
169 |
+
list_file_path = []
|
170 |
+
for idx, file in enumerate(file_obj):
|
171 |
+
file_path = file_obj.name
|
172 |
+
list_file_path.append(file_path)
|
173 |
+
# print(file_path)
|
174 |
+
# initialize_database(file_path, progress)
|
175 |
+
return list_file_path
|
176 |
+
|
177 |
+
|
178 |
+
def demo():
|
179 |
+
with gr.Blocks(theme="base") as demo:
|
180 |
+
vector_db = gr.Variable()
|
181 |
+
# qa_chain = gr.Variable()
|
182 |
+
|
183 |
+
gr.Markdown(
|
184 |
+
"""<center><h2> Document-based chatbot</center></h2>
|
185 |
+
<h3>Ask any questions about your PDF documents (single or multiple)</h3>
|
186 |
+
<i>Note: chatbot performs question-answering using Langchain and LLMs</i>
|
187 |
+
""")
|
188 |
+
with gr.Tab("Step 1 - Document pre-processing"):
|
189 |
+
with gr.Row():
|
190 |
+
document = gr.Files(height=100, file_count="multiple", file_types=["pdf"], interactive=True, label="Upload PDF Documents")
|
191 |
+
# upload_btn = gr.UploadButton("Loading document...", height=100, file_count="multiple", file_types=["pdf"], scale=1)
|
192 |
+
with gr.Row():
|
193 |
+
db_btn = gr.Radio(["ChromaDB"], label="Vector database", value = "ChromaDB", type="index", info="Choose your vector database")
|
194 |
+
with gr.Accordion("Advanced options - Text splitter", open=False):
|
195 |
+
with gr.Row():
|
196 |
+
slider_chunk_size = gr.Slider(minimum = 100, maximum = 1000, value=600, step=20, label="Chunk size", info="Chunk size", interactive=True)
|
197 |
+
with gr.Row():
|
198 |
+
slider_chunk_overlap = gr.Slider(minimum = 10, maximum = 200, value=50, step=10, label="Chunk overlap", info="Chunk overlap", interactive=True)
|
199 |
+
with gr.Row():
|
200 |
+
db_progress = gr.Textbox(label="Database Initialization", value="None")
|
201 |
+
with gr.Row():
|
202 |
+
db_btn = gr.Button("Generating vector database...")
|
203 |
+
|
204 |
+
with gr.Tab("Step 2 - Initializing QA chain"):
|
205 |
+
with gr.Row():
|
206 |
+
llm_btn = gr.Radio(["falcon-7b-instruct", "flan-t5-xxl", "mpt-7b-instruct", "Llama-2-7b-chat-hf", "Mistral-7B-Instruct-v0.1"], \
|
207 |
+
label="LLM", value = "falcon-7b-instruct", type="index", info="Choose your LLM model")
|
208 |
+
with gr.Accordion("Advanced options - LLM", open=False):
|
209 |
+
slider_temperature = gr.Slider(minimum = 0.0, maximum = 1.0, value=0.7, step=0.1, label="Temperature", info="Model temperature", interactive=True)
|
210 |
+
slider_maxtokens = gr.Slider(minimum = 256, maximum = 4096, value=1024, step=24, label="Max Tokens", info="Model max tokens", interactive=True)
|
211 |
+
slider_topk = gr.Slider(minimum = 1, maximum = 10, value=3, step=1, label="top-k samples", info="Model top-k samples", interactive=True)
|
212 |
+
with gr.Row():
|
213 |
+
llm_progress = gr.Textbox(value="None",label="QA chain Initialization")
|
214 |
+
with gr.Row():
|
215 |
+
qachain_btn = gr.Button("QA chain Initialization...")
|
216 |
+
|
217 |
+
with gr.Tab("Step 3 - Conversation"):
|
218 |
+
chatbot = gr.Chatbot(height=600)
|
219 |
+
with gr.Row():
|
220 |
+
msg = gr.Textbox(placeholder="Type message", container=True)
|
221 |
+
with gr.Row():
|
222 |
+
submit_btn = gr.Button("Submit")
|
223 |
+
clear_btn = gr.ClearButton([msg, chatbot])
|
224 |
+
|
225 |
+
# Preprocessing events
|
226 |
+
#upload_btn.upload(upload_file, inputs=[upload_btn], outputs=[document])
|
227 |
+
db_btn.click(initialize_database, inputs=[document, slider_chunk_size, slider_chunk_overlap], outputs=[vector_db, db_progress])
|
228 |
+
qachain_btn.click(initialize_LLM, inputs=[llm_btn, slider_temperature, slider_maxtokens, slider_topk, vector_db], outputs=[llm_progress]).then(lambda: None, None, chatbot, queue=False)
|
229 |
+
|
230 |
+
# Chatbot events
|
231 |
+
msg.submit(conversation, [msg, chatbot], [msg, chatbot], queue=False)
|
232 |
+
submit_btn.click(conversation, inputs=[msg, chatbot], outputs=[msg, chatbot], queue=False)
|
233 |
+
clear_btn.click(lambda: None, None, chatbot, queue=False)
|
234 |
+
demo.queue(concurrency_count=20).launch(debug=True)
|
235 |
+
|
236 |
+
|
237 |
+
if __name__ == "__main__":
|
238 |
+
demo()
|
239 |
+
|