How AI Searches by Meaning Instead of Keywords
See why searching by meaning finds answers that keyword search misses, and where keywords still win.
To show the model the right passage, you first have to find it. And finding things in text is harder than it looks, because people never phrase a question the way the document was written.
Classic search engines use keyword search: find documents that share words with the question, with extra credit for rare words that appear often in a document. It’s fast and precise, and it’s been the backbone of search for decades. But it’s literal. “My laptop won’t turn on” and “Computer fails to power up” mean the same thing and share no words at all. Keyword search shrugs and returns nothing useful.
Meaning search, also called semantic search, compares ideas instead of letters. You met the trick in Track 2: an embedding turns text into a list of numbers that captures its meaning, so texts about the same thing end up as nearby arrows. Turn the question and every article into such vectors, and the best match is simply the article whose vector points most nearly the same way, whatever words it uses. Synonyms, paraphrases and even other languages stop being a problem.
Keywords aren’t dead, though. Some queries are all about an exact string: an error code like E1043, a product number, a person’s name. Meaning search knows that “E1043” is some kind of error, but may rank every error article as roughly similar. Keyword search goes straight to the one containing those exact characters.
So each method has a blind spot the other covers. Most of this track uses meaning search because it handles how people actually ask; a later lesson shows how real systems combine both.
Key takeaways
- Keyword search matches shared words: fast and exact, but blind to synonyms and paraphrases.
- Meaning (semantic) search compares embedding vectors, so it matches ideas, not letters.
- Keyword search still wins for exact strings like codes, product numbers and names.
- The two have complementary blind spots; good systems often combine them.
Check your understanding
Go deeper (more technical)
The classic keyword ranking formula is BM25 (1990s): it scores documents by how often the query’s words appear, giving rare words more weight and stopping long documents from winning just by being long. It powers search engines such as Elasticsearch and Lucene.
Meaning search usually uses cosine similarity between embeddings (Track 2). Embedding models trained on many languages place a question in Spanish near the matching English document, a big advantage for multilingual help centres.
The simulation uses small hand-made vectors.