A method for automatically generating miniflux RSS article AI summaries using the OpenAI API

[zhu327]

Hello everyone, I've built a small tool using Rust that automatically generates AI summaries for Miniflux RSS articles using the OpenAI API. It can be deployed for free on Cloudflare Workers and utilizes Cloudflare Workers AI's LLM to generate AI summaries at no cost.

https://github.com/zhu327/miniflux-ai

[thallada]

Very cool! I think I might fork this later to play around with it.

I'm amazed that this is all free. It looks like the model is limited to 150 requests per minute so you could process a lot of articles with this too. I wonder if you could trigger the worker to process every new batch of articles with the miniflux new_entries webhook.

I'm curious why you limited the articles it summarizes to only websites in a whitelist? And, why don't you summarize articles that contain a code block? That would exclude the majority of the entries in my feed.

[zhu327]

The new_entries webhook is indeed more real-time. Maybe you can help improve it by submitting a PR. I use a whitelist because I subscribe to many types of RSS feeds, and only the articles that need quick previews require AI summaries. Have fun!

[zhu327]

The articles excluding "pre" are because the AI-generated summaries are under the "pre" tag. To avoid generating duplicates, they were excluded. If improvements are needed, one could consider using a special HTML comment tag to indicate that it has already been generated. Of course, if using the new_entries webhook, this issue might not arise.

[thallada]

I ended up forking your code and adding my own implementation that uses the new_entries webhook from Miniflux to add AI summaries to new articles. https://github.com/thallada/miniflux-ai

I was running into an issue where the cloudflare worker was getting "cancelled" while processing the miniflux webhook request and it wasn't able to save AI summaries on all of the articles. I'm still not sure exactly why this was happening, I wasn't able to reproduce it while testing while running the worker locally with npx wrangler dev. My best guess is that the AI generating request takes too long and Cloudflare cancels it for taking too long, or maybe the miniflux webhook request times out and closes the connection which causes the worker to get cancelled. My solution to this is to save all new entries into a cloudflare KV store (which is really quick) and then process all saved entries in the store every 5 minutes separately with the scheduled cron job. It seems to be working well for me so far and I believe it should still be able to stay within the worker KV free tier, so it's still completely free.