skip to content
xmarkVol. MMXXVI
docsblog
sign insign upjoin
back to blog
productsemantic-searchai

finding signal in 5,000 bookmarks

by xmark team·may 8, 2026·6 min read

on this page

  • the recall problem
  • what semantic does differently
  • how to use it
  • the tradeoffs

the recall problem

There's a number around which your saved-things collection stops being usable. For most people, it's around 500. Below that, you can browse a list and scan with your eyes. Above that, you start needing search. And the search you have is keyword search.

Keyword search wants you to remember the words. The exact words. Not the words that mean roughly the same thing. If you saved a thread about “eventual consistency” and you search “CRDTs,” keyword loses. If you saved a tweet about “the agentic loop” and you search “tool use,” keyword loses. The bookmark is there. The recall is not.

what semantic does differently

Semantic search embeds your text — every tweet, every quoted reply, every linked article excerpt, every image alt-text — into a high-dimensional space where things that mean similar things land near each other. When you ask a question, your question is embedded the same way, and the answer is whatever sits closest.

This is not magic. It's a lookup in a different coordinate system. What feels magic is the failure mode: you can ask for a thing by what it was about, and the right tweet surfaces even though it never uses the words in your question.

how to use it

Three patterns work well at the 5,000-bookmark scale:

  1. Topic recall. “what did I save about consensus algorithms?” The top results are usually the thing you half-remembered.
  2. Citation hunting. “the thread where someone explained why fly-by-wire is safer than mechanical linkage” — often you remember the shape of the argument and nothing else. Semantic finds it.
  3. Mood searching. “posts that made me reconsider my stack choice” — semantic search picks up tonal cues better than keyword.

the tradeoffs

Semantic is not always right. It will sometimes return tweets that are about the same vibe as your question but not actually responsive. xmark surfaces the top matches and a relevance score so you can scan and pick. The chat layer takes the top matches and asks Claude to answer with citations — which is the lookup pattern that tends to win at this scale.

ready to get started?

sign up in seconds. free to start, no credit card required.

get started freesign in
xmark

built by jclvsh

support

email feedbacktwitter @jclvsh

legal

privacy policyterms of servicedata deletion

© 2026 xmark. all rights reserved.