Keyword search vs asking your notes: what's the difference?

The short answer

You search your notes for "heart attack" and get nothing, even though you're sure you wrote about it. You did. The note says "myocardial infarction". That's the gap between the two kinds of search.

Keyword search, also called lexical search, finds notes that contain the words you typed. Semantic search finds notes whose meaning is close to your query, even when the words are different. Asking your notes goes one step further: a tool runs a search like that, reads the passages it finds, and writes an answer from them.

Neither kind of search is simply better. Keyword search is precise and predictable, and it's still the right tool for names, numbers and exact phrases. Semantic search is better when you don't remember the wording. Asking is useful for questions that span many notes, as long as the answer shows you where it came from.

How keyword search works

A keyword search engine keeps an index of which words appear in which notes, a bit like the index at the back of a textbook. When you search, it looks up your words and ranks the notes that contain them. Most ranking methods give more weight to rare words and to notes where your words appear often, so a search for "tachycardia" ranks a note that mentions it five times above one that mentions it once.

That makes keyword search fast, transparent and easy to trust. You can see why a result matched: the word is right there. If there's no match, the word isn't in your notes, at least not spelled that way.

The vocabulary problem

The weakness is that people describe the same thing in different words, and they do it far more than you'd expect. In a well-known study of spontaneous word choice across five application domains, two people favored the same term with probability less than 0.20 in every case (Furnas et al., 1987). The study asked people to name things for computer systems, not to write notes, but the lesson carries over: the words you search with today often aren't the words you wrote with last term.

In notes, this shows up all the time:

• Lecture notes say "myocardial infarction"; you search "heart attack".
• One meeting note says "annual plan", another says "yearly billing".
• You wrote "BP" in a hurry; you search "blood pressure".

Keyword search misses all three, unless you think of every variant yourself.

How semantic search works

Semantic search turns each passage, and your query, into a list of numbers that represents its meaning. These lists are often called embeddings. Passages with similar meanings end up with similar numbers, so "heart attack" lands close to "myocardial infarction" even though they share no words. The search returns the passages closest to your query.

On question-answering tasks, this can make a real difference. In the paper that introduced one widely used approach, the semantic retriever beat a strong keyword baseline by 9% to 19% absolute in top-20 passage retrieval accuracy across several open-domain question answering datasets (Karpukhin et al., 2020).

But it doesn't always win. A later benchmark tested retrieval systems on 18 datasets from different domains, including some they hadn't been trained for. Its authors reported that keyword search (a method called BM25) is a robust baseline, and that dense retrievers, the kind most semantic search uses, often underperform other approaches in that setting (Thakur et al., 2021). In plain terms: semantic search helps most when your question is phrased differently from your notes, and it can stumble on specialised vocabulary it hasn't seen much of.

Asking your notes: search plus an answer

When a tool answers a question from your notes, it usually works in two steps. First it searches, often with semantic search, sometimes combined with keywords. Then a language model reads the passages it found and writes an answer. This is called retrieval-augmented generation, or RAG.

The answer can only be as good as the search. If the right passage isn't found, the model may answer from the wrong one, or fill the gap with something plausible. That's why sources matter: an answer that cites the passage it used is one you can check in seconds. Our guide to why AI answers make things up covers this in more detail.

When keyword search is still the better tool

Reach for plain keyword search when:

• You know the exact word: a drug name, a person, a company, a reference number.
• You need an exact phrase, such as checking a quotation before you use it.
• You need every note that contains a term, not just the most relevant few.
• The term is unusual or technical, such as "atrioventricular node", where meaning-based search may blur it into something more common.
• You want to see exactly why a result matched.

When asking your notes works better

Asking, or semantic search, earns its place when:

• You remember the idea but not the wording.
• The answer is spread across several notes written at different times.
• You want a summary across notes, such as "what did customers say about onboarding?" or "which lectures covered blood pressure regulation?".
• You're testing yourself and want a question answered from your own material, not the open web.

A quick way to choose

1. 1If you can name the exact word or phrase, search for it.
2. 2If the search comes back empty, try two or three synonyms before giving up.
3. 3If you're asking a question rather than looking for a word, ask, and open the cited source before you rely on the answer.
4. 4If the tool can't point to where an answer came from, treat the answer as a lead, not a fact.

That last point is how Lumyo approaches asking your notes. You choose which notes a chat reads, each answer cites its sources, and tapping a source opens the exact line. When your notes don't cover a question, it says it doesn't know instead of guessing.