Hi and thanks for the positive feedback, if you use embedding you might be measuring similarity based on the face that titles or description are similar, which is not the case, because if the two items have totally different titles but they still belong to the same director (or category) then you will have low similarity score but actually there are considered similar because of the director.

Similarity is based on the context, in this context (movies) I would implement my own which combines multiple scores e,g, give a weight for each question: are they in the same category? are they by the same director? how many actors appear in both (normalize this of course? what is the cosine similarity between the two?

Then do some tuning to give reasonable weight for each metric,

I hope I answered your question.

Cheers.