Semantic Search: Finding meaning with AI

A match made in SQL

If you’re looking for something in a regular ol’ database, a keyword search could do the trick.

SELECT *
FROM dogs
WHERE breed = 'greyhound';

This example assumes a pretty basic table of dog breeds with a couple of descriptive columns. The WHERE clause is direct about what it should return: any record where the breed column is “greyhound”.

Maybe we want something that’s kind of like a greyhound. We can loosen up the WHERE to give us records where the breed is like “greyhound”. Or in other words, all records where the word “greyhound” appears in the breed column:

SELECT *
FROM dogs
WHERE breed ILIKE '%greyhound%';

And this would return an extra match:

But things aren’t always that simple. Maybe you’re looking for a dog breed that’s skinny and fast (like a greyhound). We’re not talking about keyword matches anymore. We’re talking about something more meaningful. We need to do a semantic search, and it’s something that a vector database does very well.

Finding meaning…

So, “skinny and fast” huh? Here are some results from a vector database of dog breeds, with embeddings generated using an AI model:

This seems wrong. None of these results include the word “skinny” or “fast”. Why?

Well, a vector database is able to move beyond the language of words and on to the proximity of meaning by using a text embedding model. A text embedding model can discern that the meaning of skinny and fast can be inferred from the notes of each row returned.

Using the example above, it’s able to take a numerical representation (called a text embedding) of the words “skinny and fast” and use it to find a close match of the numerical representation of the text in the notes column. Instead of matching on keywords, we can match on semantic similarity.

A quick way to see semantic similarity in action: Ask your favorite AI chat to guess one word you’re thinking of based on 5 related words or phrases. See how close it gets to the word you’re thinking. Here’s a prompt to start with.

Your job is to guess the single word I'm thinking of.
I will provide 5 related words or phrases.
Respond with one word only.