Open monatis opened 1 year ago
@monatis oh hey Monatis! long time
yea i can build this, it is straightforward
why Qdrant and not just use something like faiss
Thank you @lucidrains, kudos to you for the awesome work. Unlike Faiss's working as a library, Qdrant is a complete vector DB with support for distributed mode, filterable payload of rich data types, multiple vectors per record (say paired image and text vectors stored in the same data point) etc. I believe that it can open up more research opportunities in the retrieval-augmented paradigm. I'm also working to index LAION datasets in Qdrant and then share as a snapshot, and it can make it easier to start with experiments.
@monatis ahh i see, i think it is probably best to stick with faiss given the licensing
faiss should be good enough for use in production. many companies use it
consulted an expert, and he gave a good suggestion to build out an interface so that any of these services besides faiss could be plugged in
I'll think about it
Yeah, fair enough. We can implement support for them on top of that interface. Btw, I'm not a licensing expert but Qdrant's Apache 2.0 license should also be fine in any production use cases I guess --there are companies using it. And Apache 2.0 may even be preferable to MIT for some companies as the former explicitly grants patent rights.
@monatis ohh got it, good to know! i'm not well versed with these licenses, all i know is MIT is the most unrestrictive
@monatis This would be a great feature. I'm really interested in this. Any updates?
I would be appreciating it if Re-imagen could be implemented and integrated. Looking forward to the release.
Looking forward to the implementation.
The followup paper Re-Imagen: Retrieval-Augmented Text-to-Image Generator accesses an external memory and prepends retrieved similar entries to the input when generating. I can implement it in a PR with Qdrant as the vector DB if you would like to support it in this repo.