Skip to content

Search

Search finds exact words and similar ideas at the same time. You and the AI use the same search. Numbers never come from search: exact math tools count them.

The question “knee pain since March” first becomes exact dates, 1 March to 1 November, by Entry date. Then a search for exact words and a search for meaning run together. Their lists are merged; when two matches are equally good, the newer one wins. In AI search, a small model re-checks the top 20 and keeps the best 10. Each result comes back with the text around it and a link to its entry.Question“knee painsince March”Dates first1 Mar – 1 Novby Entry dateExact words“knee”Meaning“leg hurt on hills”Mergeties: newerre-check: 10Resultstext around+ source link
Time words become dates first. An entry that both searches find beats a strong hit in one. Recency only breaks ties.

What works today

The engine searches by words and by meaning together. Example: you ask the AI “when did my knee hurt?” and it finds “Leg sore after the hill run”, which has none of your words.

Part Today
The search box (Ctrl+K) titles and text on your device, on every key press. When you pause for a moment, results By meaning join below them: entries the word search did not show. Without a meaning model or a key, the box stays words only and nothing waits.
AI search (the search and related tools) words and meaning, merged, with a small lift for recent entries
The meaning model one for your account: baai/bge-m3
Making vectors in the desktop app on. With the local model (after your click in Settings → Plan & billing → Search by meaning): on this computer, for every collection. Without it: through OpenRouter with your own key, for Any AI collections only. With neither: words only, and AI search says so in its answer.
The re-check of the top 20 on, in AI search only: OpenRouter’s rerank with your own key (Cohere Rerank v3.5 unless you change the rerank route in Settings/ai.json). Without a key, the merged order stands and the AI is told why.
Counted results that wait for your OK yes: the count does not depend on your question, so it gives nothing away

How the desktop app makes vectors:

  • The local model is downloaded only when you click. It is about 2.3 GB. The download can resume, and every file is checked against its known fingerprint before use.
  • Until you download it, only Any AI collections get vectors, from the same model through OpenRouter. Only my local AI and No AI collections get word search only. Hidden entries are never sent.
  • Unchanged text is never sent twice. Vectors are kept in vectors.sqlite, next to the database, by model and by a fingerprint of the text. The app can rebuild this file.
  • A save never waits for a vector. A background helper makes vectors after the save.

We test this on Sam, a made-up person with 18 months of entries (see How we test the AI). With stand-in models, words alone found the right entry in the top 10 for 68 in 100 questions. Words and meaning together found 90, and with the re-check 93. The same test runs on the local model once it is downloaded, and on the real OpenRouter models with a test key.

The local model (BGE-M3)

In the desktop app, Settings → Plan & billing → Search by meaning offers one download: the local model. It is the same meaning model your account uses everywhere, so its vectors match the ones made through OpenRouter.

Question Answer
What it is BGE-M3, a free, open model that turns text into vectors. It reads more than 100 languages, so a note in Lithuanian and a question in English can still meet.
Who makes it BAAI, the Beijing Academy of Artificial Intelligence, a research lab
Where it comes from Hugging Face, a public home for AI models: huggingface.co/BAAI/bge-m3
How big it is about 2.3 GB (three files; almost all of it is the model’s weights)
Which version one fixed version, never “the latest”. Before use, every file is checked against its known fingerprint (a SHA-256 hash), so a changed or broken file is never loaded.
Where it is saved in the app data folder on your computer, in models/bge-m3. On Windows that is C:\Users\<you>\.kaizendex\app-data\models\bge-m3. The ⓘ next to Search by meaning shows your exact folder.
What it reads It runs on this computer, offline. So it may read every collection, also Only my local AI and No AI ones. Their text never leaves your computer.
Before you download it The same model runs through OpenRouter with your own key, for Any AI collections only.

How the download works:

  • Nothing downloads by itself. It starts only when you click Download.
  • You see how far it is, for example “Downloading BGE-M3 from Hugging Face · 1.1 of 2.3 GB · 49%”.
  • A stopped download resumes where it stopped, on your next click.
  • When it is done, the row says “BGE-M3 is on this computer · Ready”, and new vectors are made on your computer from then on.

The model files are not part of your data. They are not in your Kaizendex folder, and the app can download them again at any time.

How it finds things

  1. Time words become dates. “Since March” becomes 1 March to today. The AI uses today’s date and your time zone, which are always in its prompt. It searches the Entry date: when something happened, not when you typed it.
  2. Exact words find names, numbers and rare words, such as “Vilnius half marathon”.
  3. Meaning finds similar ideas in other words. “Knee pain” also finds “my leg hurt on the hills”. This uses vectors: numbers that stand for what a piece of text means.
  4. The two lists are merged. An entry that both searches find ranks above one strong hit in one search. Each entry shows only a few times, so the top results stay varied.
  5. Recency only breaks ties. When two matches are equally good, the newer one ranks first. A recent entry that does not match never beats one that does. For “the first time I…”, the AI switches recency off.
  6. In AI search, a small AI model re-checks the top 20 against your question and keeps the best 10.
  7. Each match comes back with the text around it: the heading and nearby sections of a long note, or the messages around a match in a chat.

Example: you ask “knee pain since March”. Search finds “my leg hurt on the hills” in your Morning run of 12 April. A note about knee pain from January is left out: it is before March.

Two speeds

The search box (you type) AI search (the AI calls a tool)
What it uses exact words and related at once, on every key press; meaning results join when you pause exact words, meaning and the re-check
Where it runs on your device; no AI model while you type desktop app: words and meaning on your computer (with the local model); the re-check through OpenRouter. Phones and the web: the meaning model and the re-check run through OpenRouter
Cost free a tiny amount per call, shown in the credit meter

related shows entries close to the one you have open. It uses vectors that are already on your device, so it needs no network.

For meaning results, your typed question must first become a vector:

Search for Desktop app Phones and the web
Exact words on your device, offline on your device, offline
related on your device, offline on your device, offline
Meaning, for a typed question the question becomes a vector on your computer; offline the question goes through OpenRouter to become a vector; needs the network

For this, only the question goes out, never your entries.

The re-check sends a little text. It reads the top 20 results, so it gets your question and, for each result, its collection, title and the short piece of text that matched. Only results from Any AI collections are sent, never Hidden ones; other results keep their place. No small model for the re-check runs on your computer yet, so the desktop app uses OpenRouter for it too.

Numbers come from math tools

Search finds entries. It does not count them. A total, an average or a change comes from a math tool that counts your entries exactly.

The question “How far did I run in October?” goes to the math tool aggregate: the sum of Distance in Workouts from 1 to 31 October. The answer is 84.2 km from 12 Workouts, and it shows how it was counted, with a link that opens the same view. Search finds entries; math tools count them, with the same formulas as your views and charts.Question“How far did I runin October?”Math toolSum of DistanceWorkouts · 1–31 Oct84.2 km12 Workouts · 1–31 OctSum of Distance · Open this viewSearch finds. Math tools count.The math tools use the same formulas as your views and charts, so the numbers match.
The answer shows how it counted, with a link that opens the same filtered view.

There are three math tools: one counts or adds up, one compares two periods, and one draws a trend. See AI tools.

Every answer shows its sources

  • Each claim links to the entry it came from. Each number links to how it was counted.
  • “It is not in your data” is a right answer. The AI does not make one up.
  • An empty search is not proof that something never happened. So the AI says what it searched: “I searched Workouts and Notes, 2024–2026, and found no knee injury.”
  • Some results can wait for your OK. The AI counts them instead of dropping them: “2 results in Diary need your OK”.

One meaning model for your account

Vectors from two different meaning models cannot be compared. It is like two maps drawn at different scales. So your account uses one meaning model everywhere: the desktop app, phones, the web and the agent home.

  • It is an open model. The desktop app runs it on your computer. Everywhere else, the same model runs through OpenRouter.
  • We pick it by its score on our test questions, in more than one language.
  • If a vector from a different model ever turns up, search stops with an error. It never gives wrong results in silence.
  • To change the model, the app makes new vectors in the background, next to the old ones. We compare the two on the test questions, switch, and then delete the old vectors. You keep searching the whole time.

Every entry is indexed as good text

Before an entry goes into the index, the app writes it as readable text. Each piece starts with a short header: collection, title, Entry date, tags and parent.

Workouts · Morning run · 27 Oct 2026 · Easy · Half marathon plan

So a section cut from a long note still says where it belongs.

Values are written as words, the way you would read them:

Value In the index
A GPS route 5.2 km · 31 min · pace 5:58 /km
Sleep stages 7 h 12 m asleep · deep 1 h 15 m · 4 wakes
A tag its name, such as “Easy”, not its id (the code the app uses inside)
A number with its unit: “72 kg”, not “72”

A bare “5.2” means nothing to search. “5.2 km” is found by “km”, and the meaning model knows it is a distance.

What is searched

What Example
Entries: title, values and text “Morning run”, 5.2 km, “felt great”
Page text: long notes and Block pages the section “Week 3” of “Half marathon plan”
File text the words inside a PDF lab report
Photo descriptions “a bowl of porridge with berries”
Voice transcripts a Diary entry you dictated
Your past chats “what did we decide about the marathon plan”

Each one is as private as its source. Hidden entries are never searched. Your Memory is a collection too, so the AI finds facts in it the same way. We do not add AI-written summaries of long text, unless tests show that they find things better.

Where it runs

Almost every search runs on your device. In the cloud (Pro, part of Cloud Sync; later), search runs only for agents and automations while every device is off.

Vectors are made by your account’s meaning model: in the desktop app on your computer, or through OpenRouter. On Pro, while every device is off, the agent home makes vectors for new entries through OpenRouter too. If the model is down, the job waits in a queue and runs when it is back. Word search keeps working meanwhile. End-to-end collections have their own rule, below.

The search index is as private as its source. The word index and the vectors follow the strictest privacy of the data they came from.

End-to-end collections

By default, the vectors of an end-to-end collection are made on your devices only.

You use How an end-to-end collection is searched
The desktop app Words and meaning. The desktop app makes the vectors with its own model. They sync locked, like your entries.
A phone or the web, and you also have the desktop app Words, plus meaning from the vectors your desktop made. The phone or the web never sends an entry’s text out to make vectors.
A phone or the web, with no desktop app Words only, unless you tick Let the AI route make search vectors in the collection’s AI settings. Only you can tick it, in the app. Its text then goes through OpenRouter to make them, and the app tells you so.

On phones and the web, a question you type may go through OpenRouter to become a vector. Only the question goes, never your entries. This never happens for a collection set to Only my local AI.

Example: your Diary is end-to-end. On your laptop, the desktop app makes its vectors. On your phone, “the day I felt stuck” finds the entry “Hard week at work” through those synced vectors. Only your question went through OpenRouter. The Diary’s text never left your devices unlocked.

Smaller things

  • Suggested links. When a note names a person or a project that exists, the app offers “Link to People / Tom?”. It is a suggestion: you accept or skip it. Nothing is linked behind your back.
  • Saved search scopes. A scope is a named set of collections, such as “Training” = Workouts + Sleep + Meals. A chat, an agent or a skill can search inside one scope by default, so results stay on topic.
  • Tools that are easy for an AI to use. Read tools have a short and a full answer, long results come in pages, every id has a readable name next to it, and an error says how to fix the call. See AI tools.

Why it works this way

  1. Exact words and meaning together, then a second check picks the best. Meaning alone blurs names and numbers. Words alone miss “leg hurt” when you ask for “knee pain”.
  2. Exact numbers from math tools, never guessed. An AI model that adds up 30 runs in its head makes mistakes. A formula does not.
  3. Every answer shows its sources, and “not in your data” is an honest answer. You can check any claim in one tap. An AI that must always answer makes things up.
  4. It understands time. Many questions about your life are about when: “since March”, “last summer”, “the first time”.
  5. The search box is instant and works on your device. Word search shows results on every key press, with no waiting for a server.
  6. It follows your privacy settings. The index is as private as its source, and end-to-end collections make their vectors on your devices.
  7. It is tested. Every promise on this page has a test. See How we test the AI.

Limits, said plainly

  • On phones and the web, meaning search for a typed question needs the network. Word search and related work offline everywhere.
  • With no desktop app, end-to-end collections have word search only, unless you allow the AI route to make their vectors.
  • AI search costs a tiny amount of AI credit per call.
  • The AI can still be wrong. That is why every answer shows its sources.