Subscribe to BlueSky Jetstream, compute an embedding for every post using MediaPipe Text Embedder, and compare cosine similarity to a query. Display any posts that are semantically similar enough to the query.
https://rectalogic.com/skysearch/?query=tariffs+are+tanking+the+stock+market
Static site running entirely in the users browser.
- Compute a text embedding for the users query text.
- Spawn a bunch of Web Workers based on navigator.hardwareConcurrency ready to compute text embeddings and compare their similarity to the users query embedding.
- Create a WebSocket subscribed to the BlueSky Jetstream. For each post received, send it to the first available worker. If none available, add the post to the backlog.
- A worker sends a message back indicating if the post embedding was similar enough to the users query (based on the similariy range slider the user controls). If it is, then append a BlueSky embed displaying the post.
Install deno. Then deno install and deno task dev.