Offline Notes That Find by Meaning: How On-Device Semantic Search Works

You remember saving it. The dentist sent an invoice confirmation in March and you typed something quick like "bayar ke drg Sari done". Three months later you search your notes for "dentist invoice" and get nothing. The word dentist is not in the note. The information is in the note, but the word you are searching with is not.

That gap is the problem semantic search solves. This post explains how an offline notes app can find notes by meaning, why the whole process runs on your phone, and where it still falls short.

What semantic search does

Semantic search finds notes by what they mean, not by which letters they contain. At capture time, the app runs your note through a small embedding model that turns the text into a vector, a long list of numbers that encodes meaning. At search time, your query goes through the same model. Notes whose vectors sit close to the query vector come back, even when they share zero keywords with what you typed.

Because both the note and the question land in the same "meaning space", "bayar ke drg Sari" and "dentist invoice" end up as neighbors. No tags, no filing discipline, and no exact wording required.

The pipeline, step by step

1. Embed at capture. An on-device model converts the note into a vector. No network call involved.
2. Store locally. Vectors live in a local database next to the notes themselves.
3. Embed the query. Your search phrase becomes a vector in the same space.
4. Find neighbors. The app ranks stored vectors by distance to the query vector.
5. Merge with keyword matches. Good implementations blend both: keyword search for exact names and numbers, semantic search for intent. Uteke's recall works this way, and version 0.18 added date anchors on top (memory that knows when).

Why this can run without a server

Embedding models kept shrinking while their quality held. A modern phone can run one comfortably. Uteke Mobile downloads its model once, and after that every capture and every search runs on the device. Airplane mode is a perfectly normal way to use the app.

This changes the privacy story in a structural way. Cloud assistants process your notes on a server, which means copies, permissions, and a breach surface. With on-device semantic search, your notes never cross a network, because no step in the pipeline needs one.

Keyword search vs semantic search

The two approaches are complements, not rivals. Exact strings still matter: an account number is an account number. A good mobile app runs both and blends the results, which is what Uteke's hybrid recall does (local SQLite plus semantic recall).

What it does not fix

Semantic search finds what you captured. It does not invent, and it does not read minds. If the note never got saved, no search technology will summon it. Capture speed is the other half of the product: Uteke opens straight into a new note, and voice dictation runs on device, so the friction between thought and saved note stays low.

Try it on your phone

Uteke is on Android as a free download (Google Play). The free tier covers 250 notes, 5 collections, 10 documents, and 5 voice dictations a day, with no account. Pro ($4.99/month or $49.99/year) lifts the limits and adds scheduled auto backup and full export. Over-cap notes stay readable; only new creation pauses.

The same recall engine also powers our open-source memory server for AI agents, if you want the server-side version of this idea (self-hosting guide).

Questions people ask

Does semantic search need an internet connection?

Not on Uteke Mobile. The embedding model runs on your phone, so search works in airplane mode.

Does semantic search replace keyword search?

No. It complements it. Semantic search covers intent and paraphrase, keyword search covers exact strings. The best apps blend both automatically.

Do my notes leave my phone?

No. Capture, storage, and search all happen on the device. There is no account and no cloud copy.