Updated September 17, 2026
Search bookmarks by what you remember
Semantic search finds things by meaning, not by matching keywords. Here is how it works, why it changed personal libraries, and where it actually earns its keep.
Facts you can cite
Clear product facts
Semantic bookmark search embeds each article as a vector and matches your query by meaning, not exact keywords. Backpockets: Read It Later App does this on-device with a MiniLM-class ONNX model (~90 MB) plus BM25 keyword search (hybrid). No cloud API key required for search.
The problem it solves
Keyword search assumes you remember the exact words.
Every search box you have ever used, from Google to your file system to your Gmail, is doing some version of keyword matching. It takes the words you type, looks for those exact words (or their obvious variants) in a bunch of documents, and shows you what it found.
That works well when you know what you are looking for. It fails when you only half-remember it. If you saved an article three months ago about how startup hiring is basically a market for lemons, and today you type "founders picking bad employees," a keyword search returns nothing helpful. The words don't overlap, even though the meaning is the same.
This is the gap semantic search closes. It looks for meaning overlap, not word overlap.
How it works
Embeddings, in one paragraph.
The core trick is called an embedding. An embedding is a way of turning any piece of text into a list of numbers (usually a few hundred of them) that captures what the text is about. Two pieces of text with similar meaning end up with similar number lists, even if they share no words.
Semantic search uses this. When you save an article, the system computes its embedding once and stores it. When you type a query, it computes the query's embedding and compares it to every stored one. The closest matches are returned as your results.
You don't see any of this. You just see: type a sentence, get back the articles that match its meaning.
Text becomes numbers
Every saved article is passed through an embedding model that turns its content into a fingerprint of a few hundred numbers.
Query becomes numbers
When you search, your query goes through the same model. You get a fingerprint too.
Closest wins
The system compares your fingerprint to every stored one and returns the closest matches by numeric distance.
Why it matters for you
The kind of question you can finally ask.
Keyword search reshapes the questions you allow yourself to ask. Because you know the search box needs exact words, you spend cognitive effort translating your fuzzy memory into terms the machine will match. Half the time you give up and Google instead.
Semantic search removes that translation step. You ask your library the way you would ask a friend who read the same article. "That thing about hiring being unpredictable" is enough. "The one about staying focused for four hours" works. "What was that piece where somebody said meetings destroy the morning" retrieves the meetings-tax essay.
The change is small on paper but large in practice: you actually go back to look for things you saved, because it stops feeling futile.
Trade-offs
Where semantic search is not the right tool.
Semantic search is not free. Two honest trade-offs:
It is worse at exact matches. If you know the URL, or the product name, or the invoice number, keyword search is faster and more precise. Semantic search is for the case where you remember the shape of the thing, not the exact identifier.
It costs compute. Every saved article has to be embedded once. On a Mac this is fast (thousands per minute for small models), but not free. Search itself is cheap because comparing fingerprints is a math operation, not a text scan.
The reasonable pattern is to have both. Keyword search for exact lookups. Semantic search for everything else.
A concrete example
Backpockets: Read It Later App running semantic search on saved articles.
Backpockets: Read It Later App is a Mac app built entirely around this idea for one specific use case: your saved reading pile. It imports Pocket, Raindrop, Safari, and browser bookmark exports. For every article, it stores the readable text and computes a local embedding using a small ONNX model that runs on your Mac.
Type a sentence in the search box and it returns articles ranked by meaning distance. The whole thing works offline. Nothing is sent to a server, no account is needed, and the entire library sits in a SQLite file on your Mac.
If you have ever felt the specific pain of knowing you saved something and not being able to find it again, this is the shape of software that solves it.
FAQ
What is semantic search in simple terms?
It finds documents by their meaning rather than by exact word overlap with your query. It converts both your query and every document into numeric fingerprints that represent meaning, then ranks documents by how close their fingerprint is to yours.
How is semantic search different from keyword search?
Keyword search matches text overlap: your query words have to appear in the document. Semantic search matches concepts. "Founders burning out" can retrieve an article titled "The Quiet Cost of Building" even when they share no words, because their meanings are close.
Does semantic search require an internet connection?
It depends on where the embedding model runs. Most SaaS tools call a cloud API. Local implementations, like Backpockets: Read It Later App, ship a small embedding model that runs on-device, so no network is needed for search.
Is semantic search always better?
No. For exact-name lookups (a specific URL, a specific product SKU) keyword search is faster and more precise. Semantic search wins whenever the searcher remembers the meaning of what they want but not the exact words.
Where can I try semantic search on my own bookmarks?
Backpockets: Read It Later App is a Mac app that runs semantic search locally over your bookmarks and saved articles. Free to try, $29 once, no account required.
Try it on your own pile
Semantic search over the things you already saved.
Free to try. $29 if you want to keep it. Local Mac app, no account.
Related
Related