Spaces:
Sleeping
Sleeping
Commit
•
91d7875
1
Parent(s):
e8603ed
Systems Ready!
Browse files- .gitignore +5 -0
- SOURCE_DOCUMENTS/constitution.pdf +0 -0
- app.py +304 -0
- constants.py +18 -0
- ingest.py +78 -0
- requirements.txt +23 -0
- run_localGPT.py +108 -0
- utils.py +37 -0
.gitignore
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
/db
|
2 |
+
/DB
|
3 |
+
/venv
|
4 |
+
/.idea
|
5 |
+
/__pycache__/
|
SOURCE_DOCUMENTS/constitution.pdf
ADDED
Binary file (414 kB). View file
|
|
app.py
ADDED
@@ -0,0 +1,304 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import time
|
3 |
+
from pathlib import Path
|
4 |
+
from textwrap import dedent
|
5 |
+
from types import SimpleNamespace
|
6 |
+
|
7 |
+
import gradio as gr
|
8 |
+
from charset_normalizer import detect
|
9 |
+
from chromadb.config import Settings
|
10 |
+
from epub2txt import epub2txt
|
11 |
+
from langchain.chains import RetrievalQA
|
12 |
+
from langchain.docstore.document import Document
|
13 |
+
from langchain.document_loaders import (
|
14 |
+
CSVLoader,
|
15 |
+
Docx2txtLoader,
|
16 |
+
PDFMinerLoader,
|
17 |
+
TextLoader,
|
18 |
+
)
|
19 |
+
|
20 |
+
# from constants import CHROMA_SETTINGS, SOURCE_DIRECTORY, PERSIST_DIRECTORY
|
21 |
+
from langchain.embeddings import HuggingFaceInstructEmbeddings
|
22 |
+
from langchain.llms import HuggingFacePipeline
|
23 |
+
from langchain.text_splitter import (
|
24 |
+
RecursiveCharacterTextSplitter,
|
25 |
+
)
|
26 |
+
import torch
|
27 |
+
|
28 |
+
# FAISS instead of PineCone
|
29 |
+
from langchain.vectorstores import Chroma
|
30 |
+
from loguru import logger
|
31 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
|
32 |
+
import argparse
|
33 |
+
|
34 |
+
parser = argparse.ArgumentParser('LocalGPT falcon', add_help=False)
|
35 |
+
parser.add_argument('--device_type', type=str, default="cuda", choices=["cpu", "mps", "cuda"], help='device type', )
|
36 |
+
args = parser.parse_args()
|
37 |
+
|
38 |
+
|
39 |
+
ROOT_DIRECTORY = Path(__file__).parent
|
40 |
+
PERSIST_DIRECTORY = f"{ROOT_DIRECTORY}/db"
|
41 |
+
|
42 |
+
# Define the Chroma settings
|
43 |
+
CHROMA_SETTINGS = Settings(
|
44 |
+
chroma_db_impl="duckdb+parquet",
|
45 |
+
persist_directory=PERSIST_DIRECTORY,
|
46 |
+
anonymized_telemetry=False,
|
47 |
+
)
|
48 |
+
ns = SimpleNamespace(qa=None)
|
49 |
+
|
50 |
+
# INSTRUCTORS_EMBEDDINGS_MODEL = "hkunlp/instructor-xl"
|
51 |
+
INSTRUCTORS_EMBEDDINGS_MODEL = "hkunlp/instructor-large"
|
52 |
+
# INSTRUCTORS_EMBEDDINGS_MODEL = "hkunlp/instructor-large"
|
53 |
+
# INSTRUCTORS_EMBEDDINGS_MODEL = "hkunlp/instructor-base"
|
54 |
+
|
55 |
+
def load_single_document(file_path: str or Path) -> Document:
|
56 |
+
"""ingest.py"""
|
57 |
+
# Loads a single document from a file path
|
58 |
+
# encoding = detect(open(file_path, "rb").read()).get("encoding", "utf-8")
|
59 |
+
encoding = detect(Path(file_path).read_bytes()).get("encoding", "utf-8")
|
60 |
+
if file_path.endswith(".txt"):
|
61 |
+
if encoding is None:
|
62 |
+
logger.warning(
|
63 |
+
f" {file_path}'s encoding is None "
|
64 |
+
"Something is fishy, return empty str "
|
65 |
+
)
|
66 |
+
return Document(page_content="", metadata={"source": file_path})
|
67 |
+
|
68 |
+
try:
|
69 |
+
loader = TextLoader(file_path, encoding=encoding)
|
70 |
+
except Exception as exc:
|
71 |
+
logger.warning(f" {exc}, return dummy ")
|
72 |
+
return Document(page_content="", metadata={"source": file_path})
|
73 |
+
|
74 |
+
elif file_path.endswith(".pdf"):
|
75 |
+
loader = PDFMinerLoader(file_path)
|
76 |
+
elif file_path.endswith(".csv"):
|
77 |
+
loader = CSVLoader(file_path)
|
78 |
+
elif Path(file_path).suffix in [".docx"]:
|
79 |
+
try:
|
80 |
+
loader = Docx2txtLoader(file_path)
|
81 |
+
except Exception as exc:
|
82 |
+
logger.error(f" {file_path} errors: {exc}")
|
83 |
+
return Document(page_content="", metadata={"source": file_path})
|
84 |
+
elif Path(file_path).suffix in [".epub"]: # for epub? epub2txt unstructured
|
85 |
+
try:
|
86 |
+
_ = epub2txt(file_path)
|
87 |
+
except Exception as exc:
|
88 |
+
logger.error(f" {file_path} errors: {exc}")
|
89 |
+
return Document(page_content="", metadata={"source": file_path})
|
90 |
+
return Document(page_content=_, metadata={"source": file_path})
|
91 |
+
else:
|
92 |
+
if encoding is None:
|
93 |
+
logger.warning(
|
94 |
+
f" {file_path}'s encoding is None "
|
95 |
+
"Likely binary files, return empty str "
|
96 |
+
)
|
97 |
+
return Document(page_content="", metadata={"source": file_path})
|
98 |
+
try:
|
99 |
+
loader = TextLoader(file_path)
|
100 |
+
except Exception as exc:
|
101 |
+
logger.error(f" {exc}, returnning empty string")
|
102 |
+
return Document(page_content="", metadata={"source": file_path})
|
103 |
+
|
104 |
+
return loader.load()[0]
|
105 |
+
|
106 |
+
def greet(name):
|
107 |
+
"""Test."""
|
108 |
+
logger.debug(f" name: [{name}] ")
|
109 |
+
return "Hello " + name + "!!"
|
110 |
+
|
111 |
+
|
112 |
+
def upload_files(files):
|
113 |
+
"""Upload files."""
|
114 |
+
try:
|
115 |
+
file_paths = [file.name for file in files]
|
116 |
+
except:
|
117 |
+
file_paths = [files]
|
118 |
+
logger.info(file_paths)
|
119 |
+
|
120 |
+
res = ingest(file_paths)
|
121 |
+
logger.info("Processed:\n{res}")
|
122 |
+
del res
|
123 |
+
|
124 |
+
ns.qa = load_qa()
|
125 |
+
|
126 |
+
return file_paths
|
127 |
+
|
128 |
+
|
129 |
+
def ingest(
|
130 |
+
file_paths: list
|
131 |
+
):
|
132 |
+
"""Gen Chroma db.
|
133 |
+
torch.cuda.is_available()
|
134 |
+
file_paths =
|
135 |
+
[]
|
136 |
+
"""
|
137 |
+
logger.info("Doing ingest...")
|
138 |
+
|
139 |
+
documents = []
|
140 |
+
for file_path in file_paths:
|
141 |
+
documents.append(load_single_document(f"{file_path}"))
|
142 |
+
|
143 |
+
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=200)
|
144 |
+
texts = text_splitter.split_documents(documents)
|
145 |
+
|
146 |
+
logger.info(f"Loaded {len(documents)} documents ")
|
147 |
+
logger.info(f"Split into {len(texts)} chunks of text")
|
148 |
+
|
149 |
+
# Create embeddings
|
150 |
+
logger.info(f"Load InstructEmbeddings model: {INSTRUCTORS_EMBEDDINGS_MODEL}")
|
151 |
+
embeddings = HuggingFaceInstructEmbeddings(
|
152 |
+
model_name=INSTRUCTORS_EMBEDDINGS_MODEL, model_kwargs={"device": args.device_type}
|
153 |
+
)
|
154 |
+
|
155 |
+
db = Chroma.from_documents(
|
156 |
+
texts,
|
157 |
+
embeddings,
|
158 |
+
persist_directory=PERSIST_DIRECTORY,
|
159 |
+
client_settings=CHROMA_SETTINGS,
|
160 |
+
)
|
161 |
+
db.persist()
|
162 |
+
db = None
|
163 |
+
logger.info("Done ingest")
|
164 |
+
|
165 |
+
return [
|
166 |
+
[Path(doc.metadata.get("source")).name, len(doc.page_content)]
|
167 |
+
for doc in documents
|
168 |
+
]
|
169 |
+
|
170 |
+
|
171 |
+
# https://huggingface.co/tiiuae/falcon-7b-instruct
|
172 |
+
def gen_local_llm():
|
173 |
+
"""Gen a local llm.
|
174 |
+
localgpt run_localgpt
|
175 |
+
"""
|
176 |
+
model = "tiiuae/falcon-7b-instruct"
|
177 |
+
|
178 |
+
if args.device_type == "cuda":
|
179 |
+
tokenizer = AutoTokenizer.from_pretrained(model)
|
180 |
+
else: # cpu
|
181 |
+
tokenizer=AutoTokenizer.from_pretrained(model)
|
182 |
+
model=AutoModelForCausalLM.from_pretrained(model, trust_remote_code=True)
|
183 |
+
|
184 |
+
pipe = pipeline(
|
185 |
+
"text-generation",
|
186 |
+
model=model,
|
187 |
+
tokenizer=tokenizer,
|
188 |
+
torch_dtype=torch.float32 if args.device_type =="cpu" else torch.bfloat16,
|
189 |
+
trust_remote_code=True,
|
190 |
+
device_map="cpu" if args.device_type =="cpu" else "auto",
|
191 |
+
max_length=2048,
|
192 |
+
temperature=0,
|
193 |
+
top_p=0.95,
|
194 |
+
top_k=10,
|
195 |
+
repetition_penalty=1.15,
|
196 |
+
num_return_sequences=1,
|
197 |
+
pad_token_id=tokenizer.eos_token_id
|
198 |
+
)
|
199 |
+
|
200 |
+
local_llm = HuggingFacePipeline(pipeline=pipe)
|
201 |
+
|
202 |
+
return local_llm
|
203 |
+
|
204 |
+
|
205 |
+
def load_qa():
|
206 |
+
"""Gen qa."""
|
207 |
+
logger.info("Doing qa")
|
208 |
+
|
209 |
+
embeddings = HuggingFaceInstructEmbeddings(
|
210 |
+
model_name=INSTRUCTORS_EMBEDDINGS_MODEL, model_kwargs={"device": args.device_type}
|
211 |
+
)
|
212 |
+
# xl 4.96G, large 3.5G,
|
213 |
+
db = Chroma(
|
214 |
+
persist_directory=PERSIST_DIRECTORY,
|
215 |
+
embedding_function=embeddings,
|
216 |
+
client_settings=CHROMA_SETTINGS,
|
217 |
+
)
|
218 |
+
retriever = db.as_retriever()
|
219 |
+
|
220 |
+
llm = gen_local_llm() # "tiiuae/falcon-7b-instruct"
|
221 |
+
|
222 |
+
qa = RetrievalQA.from_chain_type(
|
223 |
+
llm=llm, chain_type="stuff", retriever=retriever, return_source_documents=True
|
224 |
+
)
|
225 |
+
|
226 |
+
logger.info("Done qa")
|
227 |
+
|
228 |
+
return qa
|
229 |
+
|
230 |
+
|
231 |
+
def main1():
|
232 |
+
"""Lump codes"""
|
233 |
+
with gr.Blocks() as demo:
|
234 |
+
iface = gr.Interface(fn=greet, inputs="text", outputs="text")
|
235 |
+
iface.launch()
|
236 |
+
|
237 |
+
demo.launch()
|
238 |
+
|
239 |
+
|
240 |
+
def main():
|
241 |
+
"""Do blocks."""
|
242 |
+
logger.info(f"ROOT_DIRECTORY: {ROOT_DIRECTORY}")
|
243 |
+
|
244 |
+
with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
245 |
+
with gr.Accordion("Info", open=False):
|
246 |
+
_ = """
|
247 |
+
Talk to your docs (.pdf, .docx, .csv, .txt .md). It
|
248 |
+
takes quite a while to ingest docs (10-30 min. depending
|
249 |
+
on net, RAM, CPU etc.).
|
250 |
+
"""
|
251 |
+
gr.Markdown(dedent(_))
|
252 |
+
title = """
|
253 |
+
<div style="text-align: center;">
|
254 |
+
<h1>LocalGPT with Falcon</h1>
|
255 |
+
<p style="text-align: center;">Upload your docs (.pdf, .docx, .csv, .txt .md) by clicking the "Load docs to LangChain" and wait until the upload is complete, <br />
|
256 |
+
when everything is ready, you can start asking questions about the docs <br />
|
257 |
+
</div>
|
258 |
+
"""
|
259 |
+
gr.HTML(title)
|
260 |
+
with gr.Tab("Upload files"):
|
261 |
+
# Upload files and generate embeddings database
|
262 |
+
file_output = gr.File()
|
263 |
+
upload_button = gr.UploadButton(
|
264 |
+
"Load docs to LangChain",
|
265 |
+
file_count="multiple",
|
266 |
+
)
|
267 |
+
upload_button.upload(upload_files, upload_button, file_output)
|
268 |
+
|
269 |
+
chatbot = gr.Chatbot()
|
270 |
+
msg = gr.Textbox(label="Query")
|
271 |
+
clear = gr.Button("Clear")
|
272 |
+
|
273 |
+
def respond(message, chat_history):
|
274 |
+
if ns.qa is None: # no files processed yet
|
275 |
+
bot_message = "Provide some file(s) for processsing first."
|
276 |
+
chat_history.append((message, bot_message))
|
277 |
+
return "", chat_history
|
278 |
+
try:
|
279 |
+
res = ns.qa(message)
|
280 |
+
answer, docs = res["result"], res["source_documents"]
|
281 |
+
bot_message = f"{answer}"
|
282 |
+
except Exception as exc:
|
283 |
+
logger.error(exc)
|
284 |
+
bot_message = f"bummer! {exc}"
|
285 |
+
|
286 |
+
chat_history.append((message, bot_message))
|
287 |
+
|
288 |
+
return "", chat_history
|
289 |
+
|
290 |
+
msg.submit(respond, [msg, chatbot], [msg, chatbot])
|
291 |
+
clear.click(lambda: None, None, chatbot, queue=False)
|
292 |
+
|
293 |
+
try:
|
294 |
+
from google import colab
|
295 |
+
|
296 |
+
share = True # start share when in colab
|
297 |
+
except Exception:
|
298 |
+
share = False
|
299 |
+
|
300 |
+
demo.launch(share=share)
|
301 |
+
|
302 |
+
|
303 |
+
if __name__ == "__main__":
|
304 |
+
main()
|
constants.py
ADDED
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
# from dotenv import load_dotenv
|
3 |
+
from chromadb.config import Settings
|
4 |
+
|
5 |
+
# load_dotenv()
|
6 |
+
ROOT_DIRECTORY = os.path.dirname(os.path.realpath(__file__))
|
7 |
+
|
8 |
+
# Define the folder for storing database
|
9 |
+
SOURCE_DIRECTORY = f"{ROOT_DIRECTORY}/SOURCE_DOCUMENTS"
|
10 |
+
|
11 |
+
PERSIST_DIRECTORY = f"{ROOT_DIRECTORY}/DB"
|
12 |
+
|
13 |
+
# Define the Chroma settings
|
14 |
+
CHROMA_SETTINGS = Settings(
|
15 |
+
chroma_db_impl='duckdb+parquet',
|
16 |
+
persist_directory=PERSIST_DIRECTORY,
|
17 |
+
anonymized_telemetry=False
|
18 |
+
)
|
ingest.py
ADDED
@@ -0,0 +1,78 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import click
|
3 |
+
from typing import List
|
4 |
+
from utils import xlxs_to_csv
|
5 |
+
from langchain.document_loaders import TextLoader, PDFMinerLoader, CSVLoader
|
6 |
+
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
7 |
+
from langchain.vectorstores import Chroma
|
8 |
+
from langchain.docstore.document import Document
|
9 |
+
from constants import CHROMA_SETTINGS, SOURCE_DIRECTORY, PERSIST_DIRECTORY
|
10 |
+
from langchain.embeddings import HuggingFaceInstructEmbeddings
|
11 |
+
|
12 |
+
|
13 |
+
def load_single_document(file_path: str) -> Document:
|
14 |
+
# Loads a single document from a file path
|
15 |
+
if file_path.endswith(".txt"):
|
16 |
+
loader = TextLoader(file_path, encoding="utf8")
|
17 |
+
elif file_path.endswith(".pdf"):
|
18 |
+
loader = PDFMinerLoader(file_path)
|
19 |
+
elif file_path.endswith(".csv"):
|
20 |
+
loader = CSVLoader(file_path)
|
21 |
+
return loader.load()[0]
|
22 |
+
|
23 |
+
|
24 |
+
def load_documents(source_dir: str) -> List[Document]:
|
25 |
+
# Loads all documents from source documents directory
|
26 |
+
all_files = os.listdir(source_dir)
|
27 |
+
docs = []
|
28 |
+
for file_path in all_files:
|
29 |
+
if file_path[-4:] == 'xlsx':
|
30 |
+
for doc in xlxs_to_csv(f"{source_dir}/{file_path}"):
|
31 |
+
docs.append(load_single_document(doc))
|
32 |
+
elif file_path[-4:] in ['.txt', '.pdf', '.csv']:
|
33 |
+
docs.append(load_single_document(f"{source_dir}/{file_path}"))
|
34 |
+
return docs
|
35 |
+
# return [load_single_document(f"{source_dir}/{file_path}") for file_path in all_files if
|
36 |
+
# file_path[-4:] in ['.txt', '.pdf', '.csv']]
|
37 |
+
|
38 |
+
|
39 |
+
# @click.command()
|
40 |
+
# @click.option('--device_type', default='gpu', help='device to run on, select gpu or cpu')
|
41 |
+
# def main(device_type, ):
|
42 |
+
# # load the instructorEmbeddings
|
43 |
+
# if device_type in ['cpu', 'CPU']:
|
44 |
+
# device='cpu'
|
45 |
+
# else:
|
46 |
+
# device='cuda'
|
47 |
+
|
48 |
+
|
49 |
+
@click.command()
|
50 |
+
@click.option('--device_type', default='cuda', help='device to run on, select gpu, cpu or mps')
|
51 |
+
def main(device_type, ):
|
52 |
+
# load the instructorEmbeddings
|
53 |
+
if device_type in ['cpu', 'CPU']:
|
54 |
+
device='cpu'
|
55 |
+
elif device_type in ['mps', 'MPS']:
|
56 |
+
device='mps'
|
57 |
+
else:
|
58 |
+
device='cuda'
|
59 |
+
|
60 |
+
# Load documents and split in chunks
|
61 |
+
print(f"Loading documents from {SOURCE_DIRECTORY}")
|
62 |
+
documents = load_documents(SOURCE_DIRECTORY)
|
63 |
+
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=200)
|
64 |
+
texts = text_splitter.split_documents(documents)
|
65 |
+
print(f"Loaded {len(documents)} documents from {SOURCE_DIRECTORY}")
|
66 |
+
print(f"Split into {len(texts)} chunks of text")
|
67 |
+
|
68 |
+
# Create embeddings
|
69 |
+
embeddings = HuggingFaceInstructEmbeddings(model_name="hkunlp/instructor-base",
|
70 |
+
model_kwargs={"device": device})
|
71 |
+
|
72 |
+
db = Chroma.from_documents(texts, embeddings, persist_directory=PERSIST_DIRECTORY, client_settings=CHROMA_SETTINGS)
|
73 |
+
db.persist()
|
74 |
+
db = None
|
75 |
+
|
76 |
+
|
77 |
+
if __name__ == "__main__":
|
78 |
+
main()
|
requirements.txt
ADDED
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
langchain==0.0.166
|
2 |
+
chromadb==0.3.22
|
3 |
+
llama-cpp-python==0.1.48
|
4 |
+
urllib3==1.26.6
|
5 |
+
pdfminer.six==20221105
|
6 |
+
InstructorEmbedding
|
7 |
+
sentence-transformers
|
8 |
+
faiss-cpu
|
9 |
+
huggingface_hub
|
10 |
+
transformers
|
11 |
+
protobuf==3.20.0
|
12 |
+
accelerate
|
13 |
+
bitsandbytes
|
14 |
+
click
|
15 |
+
openpyxl
|
16 |
+
einops
|
17 |
+
xformers
|
18 |
+
gradio
|
19 |
+
charset-normalizer
|
20 |
+
PyPDF2
|
21 |
+
epub2txt
|
22 |
+
docx2txt
|
23 |
+
loguru
|
run_localGPT.py
ADDED
@@ -0,0 +1,108 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from langchain.chains import RetrievalQA
|
2 |
+
# from langchain.callbacks.streaming_stdout import StreamingStdOutCallbackHandler
|
3 |
+
from langchain.vectorstores import Chroma
|
4 |
+
from langchain.embeddings import HuggingFaceInstructEmbeddings
|
5 |
+
from langchain.llms import HuggingFacePipeline
|
6 |
+
from constants import CHROMA_SETTINGS, PERSIST_DIRECTORY
|
7 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
|
8 |
+
import click
|
9 |
+
import torch
|
10 |
+
from constants import CHROMA_SETTINGS
|
11 |
+
|
12 |
+
|
13 |
+
def load_model(device):
|
14 |
+
"""
|
15 |
+
Select a model on huggingface.
|
16 |
+
If you are running this for the first time, it will download a model for you.
|
17 |
+
subsequent runs will use the model from the disk.
|
18 |
+
"""
|
19 |
+
model = "tiiuae/falcon-7b-instruct"
|
20 |
+
|
21 |
+
if device == "cuda":
|
22 |
+
tokenizer = AutoTokenizer.from_pretrained(model)
|
23 |
+
else: # cpu
|
24 |
+
tokenizer=AutoTokenizer.from_pretrained(model)
|
25 |
+
model=AutoModelForCausalLM.from_pretrained(model, trust_remote_code=True)
|
26 |
+
|
27 |
+
pipe = pipeline(
|
28 |
+
"text-generation",
|
29 |
+
model=model,
|
30 |
+
tokenizer=tokenizer,
|
31 |
+
torch_dtype=torch.float32 if device =="cpu" else torch.bfloat16,
|
32 |
+
trust_remote_code=True,
|
33 |
+
device_map=device if device =="cpu" else "auto",
|
34 |
+
max_length=2048,
|
35 |
+
temperature=0,
|
36 |
+
top_p=0.95,
|
37 |
+
top_k=10,
|
38 |
+
repetition_penalty=1.15,
|
39 |
+
num_return_sequences=1,
|
40 |
+
pad_token_id=tokenizer.eos_token_id
|
41 |
+
)
|
42 |
+
|
43 |
+
local_llm = HuggingFacePipeline(pipeline=pipe)
|
44 |
+
|
45 |
+
return local_llm
|
46 |
+
|
47 |
+
|
48 |
+
# @click.command()
|
49 |
+
# @click.option('--device_type', default='gpu', help='device to run on, select gpu or cpu')
|
50 |
+
# def main(device_type, ):
|
51 |
+
# # load the instructorEmbeddings
|
52 |
+
# if device_type in ['cpu', 'CPU']:
|
53 |
+
# device='cpu'
|
54 |
+
# else:
|
55 |
+
# device='cuda'
|
56 |
+
|
57 |
+
|
58 |
+
## for M1/M2 users:
|
59 |
+
|
60 |
+
@click.command()
|
61 |
+
@click.option('--device_type', default='cuda', help='device to run on, select gpu, cpu or mps')
|
62 |
+
def main(device_type, ):
|
63 |
+
# load the instructorEmbeddings
|
64 |
+
if device_type in ['cpu', 'CPU']:
|
65 |
+
device='cpu'
|
66 |
+
elif device_type in ['mps', 'MPS']:
|
67 |
+
device='mps'
|
68 |
+
else:
|
69 |
+
device='cuda'
|
70 |
+
|
71 |
+
print(f"Running on: {device}")
|
72 |
+
|
73 |
+
embeddings = HuggingFaceInstructEmbeddings(model_name="hkunlp/instructor-base",
|
74 |
+
model_kwargs={"device": device})
|
75 |
+
# load the vectorstore
|
76 |
+
db = Chroma(persist_directory=PERSIST_DIRECTORY, embedding_function=embeddings, client_settings=CHROMA_SETTINGS)
|
77 |
+
retriever = db.as_retriever()
|
78 |
+
# Prepare the LLM
|
79 |
+
# callbacks = [StreamingStdOutCallbackHandler()]
|
80 |
+
# load the LLM for generating Natural Language responses.
|
81 |
+
llm = load_model(device)
|
82 |
+
qa = RetrievalQA.from_chain_type(llm=llm, chain_type="stuff", retriever=retriever, return_source_documents=True)
|
83 |
+
# Interactive questions and answers
|
84 |
+
while True:
|
85 |
+
query = input("\nEnter a query: ")
|
86 |
+
if query == "exit":
|
87 |
+
break
|
88 |
+
|
89 |
+
# Get the answer from the chain
|
90 |
+
res = qa(query)
|
91 |
+
answer, docs = res['result'], res['source_documents']
|
92 |
+
|
93 |
+
# Print the result
|
94 |
+
print("\n\n> Question:")
|
95 |
+
print(query)
|
96 |
+
print("\n> Answer:")
|
97 |
+
print(answer)
|
98 |
+
|
99 |
+
# Print the relevant sources used for the answer
|
100 |
+
print("----------------------------------SOURCE DOCUMENTS---------------------------")
|
101 |
+
for document in docs:
|
102 |
+
print("\n> " + document.metadata["source"] + ":")
|
103 |
+
print(document.page_content)
|
104 |
+
print("----------------------------------SOURCE DOCUMENTS---------------------------")
|
105 |
+
|
106 |
+
|
107 |
+
if __name__ == "__main__":
|
108 |
+
main()
|
utils.py
ADDED
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import openpyxl
|
2 |
+
import csv
|
3 |
+
import tempfile
|
4 |
+
|
5 |
+
|
6 |
+
def xlxs_to_csv(file_path: str, sheet_name: str = None) -> list[str]:
|
7 |
+
"""
|
8 |
+
Convert a workbook into a list of csv files
|
9 |
+
:param file_path: the path to the workbook
|
10 |
+
:param sheet_name: the name of the sheet to convert
|
11 |
+
:return: a list of temporary file names
|
12 |
+
"""
|
13 |
+
# Load the workbook and select the active worksheet
|
14 |
+
wb = openpyxl.load_workbook(file_path)
|
15 |
+
# ws = wb.active
|
16 |
+
#
|
17 |
+
# # Create a new temporary file and write the contents of the worksheet to it
|
18 |
+
# with tempfile.NamedTemporaryFile(mode='w+', newline='', delete=False) as f:
|
19 |
+
# c = csv.writer(f)
|
20 |
+
# for r in ws.rows:
|
21 |
+
# c.writerow([cell.value for cell in r])
|
22 |
+
# return f.name
|
23 |
+
# load all sheets if sheet_name is None
|
24 |
+
wb = wb if sheet_name is None else [wb[sheet_name]]
|
25 |
+
temp_file_name = []
|
26 |
+
# Iterate over the worksheets in the workbook
|
27 |
+
for ws in wb:
|
28 |
+
# Create a new temporary file and write the contents of the worksheet to it
|
29 |
+
with tempfile.NamedTemporaryFile(mode='w+', newline='', suffix='.csv', delete=False) as f:
|
30 |
+
c = csv.writer(f)
|
31 |
+
for r in ws.rows:
|
32 |
+
c.writerow([cell.value for cell in r])
|
33 |
+
temp_file_name.append(f.name)
|
34 |
+
# print(f'all Sheets are saved to temporary file {temp_file_name}')
|
35 |
+
return temp_file_name
|
36 |
+
|
37 |
+
|