A retailer wants to find descriptions with similar meaning even when they share no product keywords. Which representation best supports this requirement?

1. Embeddings from a model trained for semantic similarity
2. Hashes of normalized descriptions
3. Document byte lengths as one-dimensional vectors
4. Alphabetically sorted token identifiers