10k vectors is pretty small. you should be able to get away with packing all of your vectors into a matrix in e.g pytorch and doing a simple matrix vector product . should take ~tens of milliseconds.
He’s not complaining about slowness, otherwise he’d say something like ‘we tried FAISS and it was still too many milliseconds per lookup’. If his embeddings are bad, then even an exact nearest-neighbors lookup by brute-forcing every possible match (which, as you say, is more feasible than people usually think) won’t help. You’ll get the same bad answer.
10k vectors is pretty small. you should be able to get away with packing all of your vectors into a matrix in e.g pytorch and doing a simple matrix vector product . should take ~tens of milliseconds.
That won’t help if his embeddings are bad:
He’s not complaining about slowness, otherwise he’d say something like ‘we tried FAISS and it was still too many milliseconds per lookup’. If his embeddings are bad, then even an exact nearest-neighbors lookup by brute-forcing every possible match (which, as you say, is more feasible than people usually think) won’t help. You’ll get the same bad answer.