tinykoi
All posts

Search by meaning: finding the bookmark you cannot name

· Product · 4 min read

Semantic search for bookmarks finds pages by what they are about, not by the exact words you type. You search for "how to stop my sourdough going flat" and it can surface the bookmark titled "Why your starter is weak", even though the two share almost no words. In tinykoi it is called search by meaning, it is a Pro feature, and it is off until you switch AI on.

This post explains how it differs from ordinary search, when it helps, and what leaves your account when you use it.

Keyword search and its blind spot

Keyword search matches the letters you typed. It is fast, predictable and excellent when you remember a word that is actually on the page. tinykoi's search handles this well: it checks titles, tags and notes, and on every plan it searches every word of every saved page, showing the lines that matched.

The blind spot is vocabulary. You saved a page using one word, and months later you search with another: "car" for "automobile", "cheap" for "budget", "fix" for "troubleshoot". Or you only remember the idea: that article about why meetings feel unproductive. No shared word, no match.

How semantic search for bookmarks works

When AI is on, tinykoi turns each bookmark into a numerical fingerprint of its content, called an embedding. When you search, your query gets the same treatment, and tinykoi looks for bookmarks whose fingerprints sit close to the query's. Close means similar in meaning.

The results are added to the keyword results, in the same list. Specifically, a search returns the usual keyword matches plus up to 30 matches by meaning, sorted and paged the same way as everything else. You do not switch modes. Typing a query is enough.

Some practical limits, straight from how it is built:

  • Queries shorter than three characters skip it and use keywords only.
  • It looks across your own bookmarks only.
  • If the AI provider is slow or unavailable, search quietly falls back to keywords instead of failing. There is a two-second limit on the embedding step.
  • There is a daily cap of 1,000 meaning-searches per account.

An example

Say you saved a long post about how teams decide things without endless meetings. It is titled "Async by default" and never uses the word "productive". Months later you search for "why are my meetings pointless". Keyword search finds nothing, because none of those words are on the page. Search by meaning sees that the query and the post are about the same thing, and lists it. You still see the usual keyword matches first or alongside, so a precise search is never worse off.

When it helps

It earns its place in a few situations.

Fuzzy memory. "That thing about habits and identity." You remember the gist and a rough topic, not the title.

Mixed vocabulary. A library that spans fields where the same idea has different names. Medicine and engineering both talk about failure modes in different words.

Big libraries. The more you have saved, the more likely you used a different word from the one you search with today.

It helps less when you know a distinctive word. Searching for an error message or a function name is a job for keyword search, which is exact. That is why tinykoi shows both together rather than replacing one with the other.

What gets sent to AI providers

This is the part to read carefully. Search by meaning needs an outside embedding service, so some of your data leaves tinykoi. The privacy page is the authority, and says this:

  • Providers are OpenRouter and DeepInfra.
  • They are used only if you turn on AI on Pro.
  • For each bookmark they receive its title, site, excerpt, note, tags and the first 6,000 characters of its text.
  • They also receive your search queries.
  • This happens on zero-data-retention terms: nothing is stored or used for training.

If you leave AI off, none of that is sent. Search then works on keywords alone, with the full-text search that every plan has.

Think about what you save before turning it on. If some bookmarks are sensitive, know that their text excerpt is part of what gets embedded. That may be fine, or you may prefer to keep AI off. Both are reasonable.

Turning it on

Search by meaning is part of the AI settings on Pro. Open Settings, find the AI section, and switch it on. On the free plan that section shows an upgrade prompt instead, because the feature is Pro only.

The same switch also enables AI tagging, where bookmarks saved without tags get up to five, reusing tags you already have. You do not have to love one to use the other, but they share the setting.

Settings shows how many of your bookmarks are indexed for search so far, so you can see it catch up. Turning AI off deletes the search index, though tags it already added stay.

Does it replace organising?

No. A tidy system of collections and tags still makes browsing faster than any search, and meaning-search works better on well-titled pages. Treat it as a safety net for the bookmark you cannot find, not a reason to stop filing. Collections, tags or both is a good place to start with the filing.

Try it

You can see how keyword search feels first on the free plan, then add Pro when you want meaning too. Check pricing for what Pro includes, or sign up free, no card needed.