How semantic search works
A language model converts each title's description, genre and other metadata into a vector, a list of numbers that represents its meaning. The viewer's query is converted the same way. The search returns the titles whose vectors lie closest to the query's.
Because closeness is about meaning, heist and bank robbery land near each other, and a misspelled or loosely worded query still finds its target.
Semantic vs keyword search
The two are usually combined, so that an exact title match ranks first and meaning-based results fill in below.
- Keyword search: matches the words typed against titles, cast and tags. Exact, fast, and best when the viewer knows the name
- Semantic search: matches intent. Best when the viewer knows what kind of thing they want but not what it is called
When a catalog needs it
The benefit grows with the size and variety of the catalog. In a small library, browsing is enough. In a large one, many searches return nothing under keyword matching, and each empty result is a viewer who may leave. Semantic search turns those into plays.
Its quality depends on the metadata it reads. Thin synopses give it little to work with.
Last reviewed