Something I've noticed is that Gemini through gemini.google.com or through the mobile apps is vastly inferior to Gemini through aistudio.google.com. Much worse handling of long contexts amongst other things. Very odd that a product that is free (AI Studio use is free), is much worse than the product I am paying 20 quid a month for.
I find this to be especially true for the newer models like "gemini-2.5-pro-preview-03-25", so if you haven't tried AI Studio yet, I'd give that a go.
An approach I moved from this to, was to use the DB trigger to write jobs directly to an Oban oban_jobs table.
Oban jobs run almost instantly when they are queued so there's no perceptible difference in speed, but you get all the benefits of resilience that Oban provides, so your webapp connection can go down, such as during a deploy, and still catch up reliably.
It's also handy to be able use the oban_jobs table for debugging purposes.