Automatically fetch daily arxiv papers, filter with GPT, and send you an email.

The program will check for new papers every 6 hours, use GPT to filter papers related to your input keywords, and send an email to your email address. It also saves jsons under papers/.

1. Ensure that your network can connect to arxiv API and ChatGPT API

2. Install packages

   pip install openai arxiv markdown2

3. Save your OpenAI API key in openai_key.txt. If you don't want to use GPT filter or don't have an OpenAI API key, set gpt_filter=False when initializing ArxivAssistant.

4. Set up SMTP in your email (Instructions) and save the related information in mail_info.json. An example:

   {
       "mail_host": "smtp.qq.com", // SMTP host
       "mail_user": "[email protected]", // your email address
       "mail_pass": "xxxxxxxx" // e.g. identification code of SMTP
   }

5. Run the routine. Make sure the program runs constantly, e.g. run with tmux on a server

   import openai
   import json
   
   from assistant import ArxivAssistant
   
   with open("openai_key.txt") as f:
       openai.api_key = f.read()
       
   with open("mail_info.json") as f:
       mail_info = json.load(f)
   
   assistant = ArxivAssistant(
       mail_host=mail_info["mail_host"],
       mail_user=mail_info["mail_user"],
       mail_pass=mail_info["mail_pass"],
       
       categories=['cs.CV', 'cs.CL', 'cs.LG', 'cs.AI'], # your interested arxiv categories. See https://arxiv.org/category_taxonomy 
       keywords=['large language model', 'LLM'], # keywords describing your research interest
       negative_keywords=['medical'] # (Optional) keywords describing papers you don't want to read
   )
   
   assistant.run_routine()

1. Configure the number of papers:

   1. max_results_per_category: If the number of papers in some category of one day exceeds this number, only the first max_results_per_category papers are kept. Defaults to 500.

   2. max_papers_per_query: The papers are divided into groups to avoid exceeding the context length of GPT, each containing this number of papers. Defaults to 50.

   3. num_filtered_papers: The maximum number of output papers for each group. Defaults to 10.

2. Configure routine interval: Set routine_interval_hours. Defaults to 6.

Note: Arxiv publishes new papers at 20:00 EST every Sunday to Thursday. When the interval is less than 24, the routine only succeeds one time a day. When the interval is more than 24, only the last publish date (yesterday / last Thursday) is considered.

3. Configure GPT:

   1. temperature: output temperature. Defaults to 0.7.
   2. gpt_model: gpt model to use. Please note the context length, and change max_papers_per_query accordingly. Defaults to gpt-3.5-turbo-16k.

4. Change email receivers: mail_receivers is a list of receivers' email addresses. Defaults to the same as the mail sender.

5. Customize prompt for GPT and email content: Update the strings in prompts.py and single_paper_info in assistant.py. The email content follows Markdown format.

This repository is partially built on wbs2788/Arxiv-Daily.