Would the embeds from openai's CLIP be suitably "dense" for this to work? The usage requires calculating the cosine similarity between the embeds and there are

Would the embeds from openai's CLIP be suitably "dense" for this to work? The usage requires calculating the cosine similarity between the embeds and there are many instances where a lookup table approach is already being applied e.g. github.com/rom1504/clip-retrieval.