Hey Simon! Thanks for sharing this. I have long admired your work on Datasette :) We will check out your posts for ideas on coping with prompt injection.
I just came across your recent post on the ChatGPT SQL function in SQLite [1]. We just added a ChatGPT-based UDF in EVA [2]. I would love to hear your thoughts on the difference between these two approaches.
Another coincidence is that EVA uses SQLite for managing structured data by default. Can EVA's SQLite database be an interesting use case for Datasette?
The approaches look pretty similar. My chatgpt() function is pretty much the most basic possible implementation of that pattern - it's just a SQLite custom-SQL function written in Python.
You should absolutely try pointing Datasette at that SQLite database, I imagine it would work really well!
Thanks so much for sharing your thoughts! I also felt that they are pretty similar. But, I am guessing that SQLite (similar to most relational database systems) does not automatically cache the results of functions, do non-trivial cost-based optimization for functions in queries, or reorder function-based predicates based on the estimated cost of running the functions, etc.
You can cache function results yourself in Python if you want to - my implementation also sums up the tokens used by the calls to the functions.
Influencing optimization isn't possible using regular Python-based custom SQL functions though. I think you can influence that stuff in SQLite if you create more complex virtual table functions, but those aren't exposed through the regular Python sqlite3 module yet.
Thanks for the clarifications. Token summation is a cool optimization :)
Query optimizers in SQL database systems typically optimize based on the time to execute the function on a local server. The token summation optimization generalizes time-based optimization of local functions to dollar-based optimization for remote functions.
