This code was originally a fork of https://github.com/bluesky-social/feed-generator - and you can look there for more information about the mechanics of feed generation. In particular - for questions about 'how to publish the feed' from the script in the scripts folder.
I have put together a guide that tracks how I host the tool - you can find it at https://bossett.io/setting-up-bossetts-bluesky-feed-generator/. It walks you through setup of the app on Digital Ocean.
Watches for 🧪 posted by people on a set of watchlists (e.g. https://bsky.app/profile/did:plc:jfhpnnst6flqway4eaeqzj2a/lists/3jx3w32axax2f)
Feed at https://bsky.app/profile/did:plc:jfhpnnst6flqway4eaeqzj2a/feed/for-science
Watches for case-insensitive 'auspol', putes them in the feed and adds those posters to a list (e.g. https://bsky.app/profile/did:plc:jfhpnnst6flqway4eaeqzj2a/lists/3jzy2aybxwz2f).
Intended to aid discovery of new auspol posters, as well as provide the feed function.
Feed at https://bsky.app/profile/did:plc:jfhpnnst6flqway4eaeqzj2a/feed/auspol
A proof-of-concept feed suggested by @sweetbee.vip. Uses data from a user's bio in order to populate them into a feed.
A feed idea from @frecksandframes.bsky.social. Works like the What's Dad feed, but pulls terms related to both Neurodivergence and 18+. Check the file src/algos/18-plus-nd.ts for the matching expression to see which terms are being matched for each category.
The one that may get me in trouble. Suggested by @xed.bsky.social, this ranks posts from the last few days by number of replies (and replies-of-replies). Does not count self-replies, and does not count posts past the # of likes.
Strongly considering how this will evolve over time, but adjustments to:
- Let people opt-out (by following a specific account)
- Use the likes:replies ratio to 'discount' replies by a % to keep dogpiles out
- Adjust the timing to reduce how long things stay topical
I run this with Digital Ocean App Platform, with their MongoDB as an attached service.
You probably don't need Mongo if you're just doing something simple - and I've put all the DB work in src/db/dbClient.ts to make that easy to change.
I am deploying with Docker rather than the default Node containers. This was more important early on for control over the exact Node version, as certain dependencies were linked tightly to Node 18.
The tool is built to have each algorithm self-contained within a file in src/algos. Each algorithm should export both a handler function and manager class (that can inherit from algoManager - see the for-science feed). The manager is expected to implement filter methods (e.g. filter_post) that will match events that the algorithm will later deal with.
Where there's a match, the post will be stored in the database, tagged for the algorithm that matched. This can be used later in the handler function to identify posts that the algorithm should return.
Feeds will have periodicTask called every X minutes from the environment setting in FEEDGEN_TASK_INTEVAL_MINS - this is for things like list updates, or time consuming tasks that shouldn't happen interactively.
- TODO: Rename environment variables, etc. to make settings more generic
- TODO: Cache header in feed-generation.ts
- TODO: List for exclusion, header catching
- TODO: Pin function