Parsing Event Details from Seminar Announcement Emails
I receive a lot of seminar announcements, and getting that info into my calendar manually is tedious. So, I tried making a little GPT3 app to do it for me.
Given a copy-pasted email that contains all the necessary information, use OpenAI's fine-tunable GPT3 service to convert that to a JSON representation of a calendar event. Additionally, provide a download link for a generated ICS file to get this into Google Calendar or something else
Usually, when it fails, it does so because
- The JSON generation wasn't right, often because of extra quotes in the text.
- It got the wrong time zone, maybe because of overfitting to US-East.
A good solution to the first would be to not generate JSON at all--ICS is a much simpler/less-general format, and probably more tolerant of quotation shenanigans.
This is basically just a quick rewrite of the sample app provided by OpenAI, plus about $20 worth of fine-tuning on (so far) 35 emails. It costs me about half a cent per email to use it.
Unfortunately, this is an industry where backward-compatibility for a whole year is apparently too much to ask, and so, in the course of only a few months, this API and the fine-tunes associated with it were first marked as "legacy", and then removed entirely. So, this app is now broken. If I want to redo it, I'll certainly use LangChain and its more robust JSON generation capabilities.
Update 2024-01-26: I remade it with langchain instead.