AIIA Newsletter Agent

This is an intelligent AI-driven assistant that simplifies writing newsletters.

The workflow is as follows:

• Over the week (or the month) you collect links to interesting articles, papers, projects, etc. in a Telegram group (see new-bot) or directly in an Airtable base.
• The background worker (enricher) reads the links and enriches them with the title and summary.
• When you want to get a newsletter draft, you use the web interface to select the news you want to include in the newsletter and generate the text.
• As a last step, you should review the text and publish it on a platform of choice.

This code is tailored towards AI-related newsletters, but it can be easily adapted to other domains. You just need to edit the prompts of agents in the agents folder accordingly.

How to use it

Setup the AirTable base

Unfortunately, AirTable doesn't provide an API to create a base, so you'll have to do it manually.

1. Create a new AirTable base
2. Create an AirTable table Weekly newsletter in that base with the following fields:
   • URL (Primary, URL)
   • Type (Single select: 'post', 'paper', 'project', 'youtube', 'twitter', 'reddit')
   • Title (Single line text)
   • Summary (Long text)
   • Comment (Long text)
3. If you prefer to call your table differently, change the name in the news_bot_runner.py file.

Setup the project

1. If you plan to test it locally, create a .env file with the following content:

OPEN_AI_KEY=<your OpenAI API key>
AIRTABLE_API_KEY=<API key in Airtable>
BASE_ID=<Base ID in Airtable; to get it, click Help | API Documentation in your base>

2. The project is designed to be deployed on Heroku. Add the corresponding environment variables to your Heroku app. See the DEV_NOTES.md file for more details on how to deploy it on Heroku.

Run the background worker locally

python src/enrich_runner.py

Run the web interface locally

streamlit run src/weekly_newsletter_ui.py

Models

• Enrichment (i.e. reading and summarizing articles) uses GPT-3.5 by default. You can replace it with GPT-4 in enrich_runner.py, but it's not recommended: the quality of summarization is pretty much the same, but GPT-4 is much slower.
• Newsletter writer uses GPT-4 by default. You can replace it with GPT-3.5 in weekly_newsletter_ui.py. GPT-4 is generally better for complex propmts like the ones used for writing a draft.