Performance of dense retrieval / embeddings

[alexandergunnarson]

Does LLaVA work well for dense retrieval / embeddings? I assume not, because it's not what it's been trained on (see also subpar performance of GPT embeddings vs. that of purpose-trained embedding models), but perhaps it's an emergent feature. For instance, let's say I have a vector DB and I want to, given a text/image/text+image query, find related text/image results.

See also a possibly-related comment by @SiyuanHuang95 at #254 (comment).