Computers compare numbers, not meaning. An embedding model converts a sentence into a few hundred or a few thousand numbers, and it is trained so that the distance between two of those lists reflects how related the two sentences are. "How do I cancel my subscription" lands near "where is the unsubscribe button" even though they share almost no words.
This is what makes semantic search work, and it is the reason keyword search and embedding search fail in opposite ways. Keyword search misses a customer who used different words. Embedding search finds them, and occasionally returns something that is merely about the same topic rather than actually answering the question. Production systems usually run both and combine the results.
Related terms
RAG (retrieval-augmented generation)
Looking up relevant documents first and putting them in the prompt, so the model answers from your material instead of from memory.
Vector database
A database built to find the stored items whose embeddings are closest to a query's, quickly, across millions of rows.
Large language model (LLM)
A model trained on enormous amounts of text to predict what comes next, which turns out to be enough to write, summarise, translate and reason through many tasks.
The bench this belongs to
AI automationThe repetitive half of your week, handed to software that does not get bored. Inbox triage, follow-ups, reporting, data entry between tools that were never meant to talk.
