Embeddings model contextual similarity, which is also a flawed proxy for meaning. How good it is depends on how comprehensive your training data is in containing the meaning you're trying to capture