ArxivChatGPT

Chat with arxiv paper

similar project

1. arxiv-summary reddit
2. github-chatpaper
3. arxiv-sanity github website

setup local development environment

NOTICE

1. read openai's documentation or husisy/learning/openai/readme.md first
2. do NOT put API-key (openai, WEAVIATE) in the git history
3. access openai api key only in US-VPS

micromamba create -n arxiv
micromamba install -n arxiv -c conda-forge cython matplotlib h5py pillow protobuf scipy requests tqdm flask ipython openai python-dotenv tiktoken lxml tqdm pdfminer.six python-magic pylatexenc chardet flask-sqlalchemy flask-migrate flask-login wtforms flask-wtf email_validator waitress
micromamba activate arxiv
pip install weaviate-client

1. complete the .env file from .env.example
2. run python main_chat.py
3. (optional) update database: python main_database.py

>>> from main_chat import *
>>> chatgpt.list_arxiv(num_print=5)
>>> chatgpt.select('2304.02241')
>>> chatgpt.chat("What is the main contribution of this paper?")

计划

1. 角色：用户，矢量数据库，openai embedding model (ada-002), this repo（以下简称self）
2. 用户：阅读abstract
   • abstract已经足够精炼了，进一步summarize是不必要的，且会丢失信息，当然用户可以在下一步通过chat来summarize
   • reference是结构化数据，不易丢入到chatgpt中
3. 用户：chat来查询文章内容
   • input question text
   • question to embedding
   • find similar vector in database
   • concatenate similar paragraph with question
   • input to chatgpt

flow chart

1. 每天从arxiv下载文章列表 https://arxiv.org/list/quant-ph/recent
   • url_list(list,str)
2. 下载文章对应的LaTeX文件
   • tex_text(str)
   • 可能的错误：不存在。tar.gz文件，非.tar.gz文件格式，多个子目录
3. 解析LaTeX文件，按照section分段，较长section进一步分割
   • chunk_text_list(list,str)
   • 可能的错误：多个.tex文件，.tex文件包含\input{}，非section分段
   • 超参数，每个chunk的最大token数量
4. 调用openai-ada-002将每个chunk映射为矢量
5. price
   • about 0.01USD for converting one paper into chunk vectors
   • about 0.002USD for each question

.env environment parameters

# weaviate database api key
WEAVIATE_API_URL=https://xxx
WEAVIATE_API_KEY=yyy

OPENAI_API_KEY=sk-zzz

# sqlite3 database path
SQLITE3_DB_PATH=/path/to/arxiv.sqlite3

# arxiv data folder
ARXIV_DIRECTORY=/path/to/arxiv-folder

# how many token for each chunk: 400 is chosen without too much consideration
ARXIVGPT_MAX_TOKEN_PER_CHUNK=400

# how many token for each question: 1800 is chosen without too much consideration
ARXIVGPT_MAX_TOKEN_PER_QA=1800

# 1: save chunk-vector.npy locally, mainly for debug
# 0: not save
ARXIVGPT_SAVE_NUMPY_VECTOR=1

brainstom: 用户有哪些关心的问题

1. summarize this paper
   • not good question, it's just the abstract
2. what the key contributions of the content
   • not bad
3. the most related paper according to the references
   • bad question, no references input to chatgpt
4. what is the method of this paper
   • good question
5. what is the result of this paper
   • good question
6. TODO
7. explain the word "xxx" in this paper
8. explain the word "xxx" in command sense
9. TODO 常见的问题先问掉并存起来

TODO

1. naive website
2. website login to avoid abuse login (api key is expensive)
3. sqlite: a database to store the path
4. public /data folder to store the downloaded tex file
5. website: store user's input
6. website: collect user's feedback, like which papers should be included
7. website: how to show chating messages

arxiv folder structure

arxiv/
├── 2101.00001/ (arxivID)
│   ├── main.pdf
│   ├── main.tar.gz
│   ├── main.tex
│   ├── meta-info.json (abstract, title, author_list, subject)
│   ├── chunk-text.json (list os string)
│   ├── chunk-vector.npy (numpy,float64,(N,1536)): optional
│   └── untar/ (extract from .tar.gz)
└── 2304.01411/
    ├── main.pdf
    ├── main.tar.gz
    ├── main.tex
    ├── meta-info.json
    ├── chunk-text.json
    ├── chunk-vector.npy
    └── untar/

arxiv.sqlite3 table=paper

# database paper
pid: int (primary key)
arxivID: str
meta_info_json_path: str
pdf_path: str('' if not exist)
tex_path: str ('' if not exist)
chunk_text_json_path: str ('' if not exist)
num_chunk: int

weaviate table=Paper

chunk: text
arxiv_id: string
index: int
num_chunk: int
num_token: int
vector (ada-002)

misc

# turn on crawling process
nohup python main_crawl.py >> log/main_crawl.log 2>&1 &

# turn on website
nohup waitress-serve --listen="127.0.0.1:5001" "app:app" >> log/waitress-serve.log 2>&1 &

# turn off crawling process
ps ax | grep main_crawl #get PID
kill <PID> #replace <PID>

# turn off website
ps ax | grep waitress-serve #get PID
kill <PID> #replace <PID>