Put related things near each other
An embedding model learns representations useful for its training objective. Texts with related meanings can end up close under a chosen similarity measure. This lets software search beyond exact keyword matches.
Find the idea, not just the phrase
A search for “How do I get my money back?” could retrieve a document titled “Refund policy”, even without shared wording.
A search system embeds the question and compares it with document embeddings. It may then rerank the candidates before showing results or passing passages to a language model. Think of a map with many dimensions, rather than a literal two-dimensional diagram.
Nearby does not mean correct
Similarity does not establish truth, freshness or permission to access a document. Results depend on the model, document splitting and ranking. A related passage can still answer the wrong question.
KEEP EXPLORINGMake sense of another idea↗