Merge pull request #5 from jerpint/parse_docs
Browse files- .gitignore +137 -0
- buster/chatbot.py +4 -3
- buster/data/document_embeddings.csv +0 -0
- buster/data/{sections.pkl → documents.csv} +0 -0
- buster/docparser.py +61 -37
- requirements.txt +3 -2
.gitignore
ADDED
@@ -0,0 +1,137 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Byte-compiled / optimized / DLL files
|
2 |
+
__pycache__/
|
3 |
+
*.py[cod]
|
4 |
+
*$py.class
|
5 |
+
|
6 |
+
albenchmark/data/
|
7 |
+
|
8 |
+
# Ignore notebooks by default
|
9 |
+
*.ipynb
|
10 |
+
|
11 |
+
# C extensions
|
12 |
+
*.so
|
13 |
+
|
14 |
+
# Distribution / packaging
|
15 |
+
.Python
|
16 |
+
build/
|
17 |
+
develop-eggs/
|
18 |
+
dist/
|
19 |
+
downloads/
|
20 |
+
eggs/
|
21 |
+
.eggs/
|
22 |
+
lib/
|
23 |
+
lib64/
|
24 |
+
parts/
|
25 |
+
sdist/
|
26 |
+
var/
|
27 |
+
wheels/
|
28 |
+
pip-wheel-metadata/
|
29 |
+
share/python-wheels/
|
30 |
+
*.egg-info/
|
31 |
+
.installed.cfg
|
32 |
+
*.egg
|
33 |
+
MANIFEST
|
34 |
+
|
35 |
+
# PyInstaller
|
36 |
+
# Usually these files are written by a python script from a template
|
37 |
+
# before PyInstaller builds the exe, so as to inject date/other infos into it.
|
38 |
+
*.manifest
|
39 |
+
*.spec
|
40 |
+
|
41 |
+
# Installer logs
|
42 |
+
pip-log.txt
|
43 |
+
pip-delete-this-directory.txt
|
44 |
+
|
45 |
+
# Unit test / coverage reports
|
46 |
+
htmlcov/
|
47 |
+
.tox/
|
48 |
+
.nox/
|
49 |
+
.coverage
|
50 |
+
.coverage.*
|
51 |
+
.cache
|
52 |
+
nosetests.xml
|
53 |
+
coverage.xml
|
54 |
+
*.cover
|
55 |
+
*.py,cover
|
56 |
+
.hypothesis/
|
57 |
+
.pytest_cache/
|
58 |
+
|
59 |
+
# Translations
|
60 |
+
*.mo
|
61 |
+
*.pot
|
62 |
+
|
63 |
+
# Django stuff:
|
64 |
+
*.log
|
65 |
+
local_settings.py
|
66 |
+
db.sqlite3
|
67 |
+
db.sqlite3-journal
|
68 |
+
|
69 |
+
# Flask stuff:
|
70 |
+
instance/
|
71 |
+
.webassets-cache
|
72 |
+
|
73 |
+
# Scrapy stuff:
|
74 |
+
.scrapy
|
75 |
+
|
76 |
+
# Sphinx documentation
|
77 |
+
docs/_build/
|
78 |
+
|
79 |
+
# PyBuilder
|
80 |
+
target/
|
81 |
+
|
82 |
+
# Jupyter Notebook
|
83 |
+
.ipynb_checkpoints
|
84 |
+
|
85 |
+
# IPython
|
86 |
+
profile_default/
|
87 |
+
ipython_config.py
|
88 |
+
|
89 |
+
# pyenv
|
90 |
+
.python-version
|
91 |
+
|
92 |
+
# pipenv
|
93 |
+
# According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
|
94 |
+
# However, in case of collaboration, if having platform-specific dependencies or dependencies
|
95 |
+
# having no cross-platform support, pipenv may install dependencies that don't work, or not
|
96 |
+
# install all needed dependencies.
|
97 |
+
#Pipfile.lock
|
98 |
+
|
99 |
+
# PEP 582; used by e.g. github.com/David-OConnor/pyflow
|
100 |
+
__pypackages__/
|
101 |
+
|
102 |
+
# Celery stuff
|
103 |
+
celerybeat-schedule
|
104 |
+
celerybeat.pid
|
105 |
+
|
106 |
+
# SageMath parsed files
|
107 |
+
*.sage.py
|
108 |
+
|
109 |
+
# Environments
|
110 |
+
.env
|
111 |
+
.venv
|
112 |
+
env/
|
113 |
+
venv/
|
114 |
+
ENV/
|
115 |
+
env.bak/
|
116 |
+
venv.bak/
|
117 |
+
|
118 |
+
# Spyder project settings
|
119 |
+
.spyderproject
|
120 |
+
.spyproject
|
121 |
+
|
122 |
+
# Rope project settings
|
123 |
+
.ropeproject
|
124 |
+
|
125 |
+
# mkdocs documentation
|
126 |
+
/site
|
127 |
+
|
128 |
+
# mypy
|
129 |
+
.mypy_cache/
|
130 |
+
.dmypy.json
|
131 |
+
dmypy.json
|
132 |
+
|
133 |
+
# Pyre type checker
|
134 |
+
.pyre/
|
135 |
+
|
136 |
+
# VSCode
|
137 |
+
.vscode/
|
buster/chatbot.py
CHANGED
@@ -1,15 +1,16 @@
|
|
1 |
import logging
|
2 |
-
import pickle
|
3 |
|
4 |
import numpy as np
|
5 |
import openai
|
6 |
import pandas as pd
|
7 |
-
from docparser import EMBEDDING_MODEL
|
8 |
from openai.embeddings_utils import cosine_similarity, get_embedding
|
9 |
|
|
|
|
|
10 |
logger = logging.getLogger(__name__)
|
11 |
logging.basicConfig(level=logging.INFO)
|
12 |
|
|
|
13 |
# search through the reviews for a specific product
|
14 |
def rank_documents(df: pd.DataFrame, query: str, top_k: int = 3) -> pd.DataFrame:
|
15 |
product_embedding = get_embedding(
|
@@ -33,7 +34,7 @@ def engineer_prompt(question: str, documents: list[str]) -> str:
|
|
33 |
def get_gpt_response(question: str, df) -> str:
|
34 |
# rank the documents, get the highest scoring doc and generate the prompt
|
35 |
candidates = rank_documents(df, query=question, top_k=1)
|
36 |
-
documents = candidates.
|
37 |
prompt = engineer_prompt(question, documents)
|
38 |
|
39 |
logger.info(f"querying GPT...")
|
|
|
1 |
import logging
|
|
|
2 |
|
3 |
import numpy as np
|
4 |
import openai
|
5 |
import pandas as pd
|
|
|
6 |
from openai.embeddings_utils import cosine_similarity, get_embedding
|
7 |
|
8 |
+
from buster.docparser import EMBEDDING_MODEL
|
9 |
+
|
10 |
logger = logging.getLogger(__name__)
|
11 |
logging.basicConfig(level=logging.INFO)
|
12 |
|
13 |
+
|
14 |
# search through the reviews for a specific product
|
15 |
def rank_documents(df: pd.DataFrame, query: str, top_k: int = 3) -> pd.DataFrame:
|
16 |
product_embedding = get_embedding(
|
|
|
34 |
def get_gpt_response(question: str, df) -> str:
|
35 |
# rank the documents, get the highest scoring doc and generate the prompt
|
36 |
candidates = rank_documents(df, query=question, top_k=1)
|
37 |
+
documents = candidates.text.to_list()
|
38 |
prompt = engineer_prompt(question, documents)
|
39 |
|
40 |
logger.info(f"querying GPT...")
|
buster/data/document_embeddings.csv
CHANGED
The diff for this file is too large to render.
See raw diff
|
|
buster/data/{sections.pkl → documents.csv}
RENAMED
Binary files a/buster/data/sections.pkl and b/buster/data/documents.csv differ
|
|
buster/docparser.py
CHANGED
@@ -1,17 +1,20 @@
|
|
1 |
import glob
|
|
|
2 |
import os
|
3 |
-
import pickle
|
4 |
|
5 |
import pandas as pd
|
6 |
import tiktoken
|
7 |
from bs4 import BeautifulSoup
|
8 |
-
from openai.embeddings_utils import
|
9 |
|
10 |
EMBEDDING_MODEL = "text-embedding-ada-002"
|
11 |
EMBEDDING_ENCODING = "cl100k_base" # this the encoding for text-embedding-ada-002
|
12 |
|
13 |
|
14 |
-
|
|
|
|
|
|
|
15 |
"""Parse all HTML files in `root_dir`, and extract all sections.
|
16 |
|
17 |
Sections are broken into subsections if they are longer than `max_section_length`.
|
@@ -19,85 +22,106 @@ def get_all_sections(root_dir: str, max_section_length: int = 3000) -> list[str]
|
|
19 |
"""
|
20 |
files = glob.glob("*.html", root_dir=root_dir)
|
21 |
|
22 |
-
|
23 |
-
|
24 |
-
# Recurse until sections are small enough
|
25 |
-
def get_all_subsections(soup, selector: str) -> list[str]:
|
26 |
-
found = soup.select(selector)
|
27 |
-
data = [x.text.split(";")[-1].strip() for x in found]
|
28 |
|
29 |
sections = []
|
30 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
if len(section) > max_section_length:
|
32 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
33 |
else:
|
34 |
sections.append(section)
|
|
|
|
|
35 |
|
36 |
-
return sections
|
37 |
|
38 |
sections = []
|
|
|
|
|
39 |
for file in files:
|
40 |
filepath = os.path.join(root_dir, file)
|
41 |
with open(filepath, "r") as file:
|
42 |
source = file.read()
|
43 |
|
44 |
soup = BeautifulSoup(source, "html.parser")
|
45 |
-
|
|
|
46 |
|
47 |
-
|
|
|
48 |
|
|
|
49 |
|
50 |
-
|
51 |
-
with open(filepath, "wb") as f:
|
52 |
-
pickle.dump(sections, f)
|
53 |
|
|
|
54 |
|
55 |
-
def read_sections(filepath: str) -> list[str]:
|
56 |
-
with open(filepath, "rb") as fp:
|
57 |
-
sections = pickle.load(fp)
|
58 |
|
59 |
-
|
|
|
60 |
|
61 |
|
62 |
-
def
|
63 |
-
|
64 |
-
|
65 |
-
with open(fname, "rb") as fp:
|
66 |
-
documents = pickle.load(fp)
|
67 |
-
df["documents"] = documents
|
68 |
-
return df
|
69 |
|
70 |
|
71 |
def compute_n_tokens(df: pd.DataFrame) -> pd.DataFrame:
|
72 |
encoding = tiktoken.get_encoding(EMBEDDING_ENCODING)
|
73 |
-
df["n_tokens"] = df.
|
74 |
return df
|
75 |
|
76 |
|
77 |
def precompute_embeddings(df: pd.DataFrame) -> pd.DataFrame:
|
78 |
-
df["embedding"] = df.
|
79 |
return df
|
80 |
|
81 |
|
82 |
def generate_embeddings(filepath: str, output_csv: str) -> pd.DataFrame:
|
83 |
# Get all documents and precompute their embeddings
|
84 |
-
df =
|
85 |
df = compute_n_tokens(df)
|
86 |
df = precompute_embeddings(df)
|
87 |
-
|
88 |
return df
|
89 |
|
90 |
|
91 |
if __name__ == "__main__":
|
92 |
root_dir = "/home/hadrien/perso/mila-docs/output/"
|
93 |
-
save_filepath =
|
94 |
|
95 |
# How to write
|
96 |
-
|
97 |
-
|
98 |
|
99 |
# How to load
|
100 |
-
|
101 |
|
102 |
-
#
|
103 |
df = generate_embeddings(filepath=save_filepath, output_csv="data/document_embeddings.csv")
|
|
|
1 |
import glob
|
2 |
+
import math
|
3 |
import os
|
|
|
4 |
|
5 |
import pandas as pd
|
6 |
import tiktoken
|
7 |
from bs4 import BeautifulSoup
|
8 |
+
from openai.embeddings_utils import get_embedding
|
9 |
|
10 |
EMBEDDING_MODEL = "text-embedding-ada-002"
|
11 |
EMBEDDING_ENCODING = "cl100k_base" # this the encoding for text-embedding-ada-002
|
12 |
|
13 |
|
14 |
+
BASE_URL = "https://docs.mila.quebec/"
|
15 |
+
|
16 |
+
|
17 |
+
def get_all_documents(root_dir: str, max_section_length: int = 3000) -> pd.DataFrame:
|
18 |
"""Parse all HTML files in `root_dir`, and extract all sections.
|
19 |
|
20 |
Sections are broken into subsections if they are longer than `max_section_length`.
|
|
|
22 |
"""
|
23 |
files = glob.glob("*.html", root_dir=root_dir)
|
24 |
|
25 |
+
def get_all_subsections(soup: BeautifulSoup) -> tuple[list[str], list[str], list[str]]:
|
26 |
+
found = soup.find_all("a", href=True, class_="headerlink")
|
|
|
|
|
|
|
|
|
27 |
|
28 |
sections = []
|
29 |
+
urls = []
|
30 |
+
names = []
|
31 |
+
for section_found in found:
|
32 |
+
section_soup = section_found.parent.parent
|
33 |
+
section_href = section_soup.find_all("a", href=True, class_="headerlink")
|
34 |
+
|
35 |
+
# If sections has subsections, keep only the part before the first subsection
|
36 |
+
if len(section_href) > 1:
|
37 |
+
section_siblings = section_soup.section.previous_siblings
|
38 |
+
section = [sibling.text for sibling in section_siblings]
|
39 |
+
section = "".join(section[::-1])[1:]
|
40 |
+
else:
|
41 |
+
section = section_soup.text[1:]
|
42 |
+
|
43 |
+
url = section_found["href"]
|
44 |
+
name = section_found.parent.text[:-1]
|
45 |
+
|
46 |
+
# If text is too long, split into chunks of equal sizes
|
47 |
if len(section) > max_section_length:
|
48 |
+
n_chunks = math.ceil(len(section) / float(max_section_length))
|
49 |
+
separator_index = math.floor(len(section) / n_chunks)
|
50 |
+
|
51 |
+
section_chunks = [section[separator_index * i : separator_index * (i + 1)] for i in range(n_chunks)]
|
52 |
+
url_chunks = [url] * n_chunks
|
53 |
+
name_chunks = [name] * n_chunks
|
54 |
+
|
55 |
+
sections.extend(section_chunks)
|
56 |
+
urls.extend(url_chunks)
|
57 |
+
names.extend(name_chunks)
|
58 |
else:
|
59 |
sections.append(section)
|
60 |
+
urls.append(url)
|
61 |
+
names.append(name)
|
62 |
|
63 |
+
return sections, urls, names
|
64 |
|
65 |
sections = []
|
66 |
+
urls = []
|
67 |
+
names = []
|
68 |
for file in files:
|
69 |
filepath = os.path.join(root_dir, file)
|
70 |
with open(filepath, "r") as file:
|
71 |
source = file.read()
|
72 |
|
73 |
soup = BeautifulSoup(source, "html.parser")
|
74 |
+
sections_file, urls_file, names_file = get_all_subsections(soup)
|
75 |
+
sections.extend(sections_file)
|
76 |
|
77 |
+
urls_file = [BASE_URL + os.path.basename(file.name) + url for url in urls_file]
|
78 |
+
urls.extend(urls_file)
|
79 |
|
80 |
+
names.extend(names_file)
|
81 |
|
82 |
+
documents_df = pd.DataFrame.from_dict({"name": names, "url": urls, "text": sections})
|
|
|
|
|
83 |
|
84 |
+
return documents_df
|
85 |
|
|
|
|
|
|
|
86 |
|
87 |
+
def write_documents(filepath: str, documents_df: pd.DataFrame):
|
88 |
+
documents_df.to_csv(filepath, index=False)
|
89 |
|
90 |
|
91 |
+
def read_documents(filepath: str) -> pd.DataFrame:
|
92 |
+
return pd.read_csv(filepath)
|
|
|
|
|
|
|
|
|
|
|
93 |
|
94 |
|
95 |
def compute_n_tokens(df: pd.DataFrame) -> pd.DataFrame:
|
96 |
encoding = tiktoken.get_encoding(EMBEDDING_ENCODING)
|
97 |
+
df["n_tokens"] = df.text.apply(lambda x: len(encoding.encode(x)))
|
98 |
return df
|
99 |
|
100 |
|
101 |
def precompute_embeddings(df: pd.DataFrame) -> pd.DataFrame:
|
102 |
+
df["embedding"] = df.text.apply(lambda x: get_embedding(x, engine=EMBEDDING_MODEL))
|
103 |
return df
|
104 |
|
105 |
|
106 |
def generate_embeddings(filepath: str, output_csv: str) -> pd.DataFrame:
|
107 |
# Get all documents and precompute their embeddings
|
108 |
+
df = read_documents(filepath)
|
109 |
df = compute_n_tokens(df)
|
110 |
df = precompute_embeddings(df)
|
111 |
+
write_documents(output_csv, df)
|
112 |
return df
|
113 |
|
114 |
|
115 |
if __name__ == "__main__":
|
116 |
root_dir = "/home/hadrien/perso/mila-docs/output/"
|
117 |
+
save_filepath = "data/documents.csv"
|
118 |
|
119 |
# How to write
|
120 |
+
documents_df = get_all_documents(root_dir)
|
121 |
+
write_documents(save_filepath, documents_df)
|
122 |
|
123 |
# How to load
|
124 |
+
documents_df = read_documents(save_filepath)
|
125 |
|
126 |
+
# precompute the document embeddings
|
127 |
df = generate_embeddings(filepath=save_filepath, output_csv="data/document_embeddings.csv")
|
requirements.txt
CHANGED
@@ -1,4 +1,5 @@
|
|
1 |
-
|
2 |
-
openai
|
3 |
numpy
|
4 |
tiktoken
|
|
|
|
|
|
1 |
+
bs4
|
|
|
2 |
numpy
|
3 |
tiktoken
|
4 |
+
openai
|
5 |
+
pandas
|