Re:Ollama: Embedding Models
by mhuffman on Tuesday April 09, 2024 @09:00AM
(#39978956)
I do not think embeddings would have too much trouble with a mult-topic document. It might have more trouble where words are phrases are used closed together in the same document but have different meanings.
[ Parent
| on HN ]
It's a big strength of embeddings like BERT and later on that they embed not just the "word" but the word plus its context so if the same word is used in two different places with two different meanings the embedding for that token ought to be different.