Bloom

Bloom started as a tool that turns an article into a navigable knowledge tree - summary nodes in the author’s voice that expand into the actual passages, with concept terms indexed and cross-linked. It’s grown into two things sharing a repo: a web app where you subscribe to feeds and “bloom” articles into a reading library, and @bloom, a Discord bot that reads an article with you a piece at a time.

How it works

• Three services. A Go API (Postgres via pgx, no ORM), a React and Vite web client, and a separate Discord bot module, all built as their own images.
• Article to tree. The core path fetches an article, extracts the text, and asks a language model to structure it as a tree of summary and passage nodes, rendered in a self-contained viewer.
• @bloom, live and steered. Mention the bot with a link and it opens a thread, posts a short read plus labeled buttons - go deeper, tangent, outside context - and branches from whichever button you tap. Old buttons never expire; clicking one branches from its node with everything since folded in.
• Disposable agent compute. The bot leans on my session platform for authenticated, concurrency-limited agent turns, and owns the durable record of each read itself so a lost session just rehydrates.
• Runs on my cluster. The web service is deployed on Kubernetes behind my own single sign-on.

Using it

You read a bloom top-down, one node at a time. Everything starts collapsed - the root is a one-sentence summary in the author’s voice, and you expand only the branches that earn your attention. A summary node opens into sub-branches; a passage node shows just its highlighted phrases collapsed, and the full paragraphs with those phrases marked when you expand it. Concept terms in the text are clickable - tap one and a popover scrolls you back to where it was first introduced. The point is reading at your own depth: skim the summaries, drill in where you’re curious, skip the rest, instead of grinding front to back.

@bloom turns that into something live. Instead of one precomputed tree, the bot reads a piece with you - it posts a short chunk plus labeled buttons (go deeper, tangent, outside context) and branches from whichever you tap. Old buttons never expire, so you can scroll back to any earlier point and take it a different direction; the thread grows into a forum-shaped tree of how you actually read it. It feels less like querying a chatbot and more like reading with someone in the room you can keep poking.

Notes

The bot is a deliberate re-think of the original: instead of precomputing one fixed analysis tree, it’s a small forum interrogating a chatbot about an article, together.