A keyword search for expense limit misses a policy document that only says spending cap, even though they mean the same thing. Semantic search compares the embeddings of the query and the documents, so it can surface the right passage even when the words do not match exactly, which matters for company knowledge that is rarely written in one consistent vocabulary.
The limit is that semantic search returns what is similar in meaning, not necessarily what is current or authorized for the person asking. A vector database built on outdated or duplicated documents will confidently return outdated or duplicated answers, so keeping the underlying source documents current matters as much as the search technology itself.