Vector DBs 🤝 Multimodality
May 25, 2024
I saw this tweet and without thinking I reacted "yikes!". And then I thought about it and I thought "yikes this is really bad for Pinecone".
I've been a *bit* of a hater on vector DBs for some time now. Not because I don't think they're useful (I use them and think they are great!), but because people see them as a cure-all for their search and retrieval problems when there's potentially better ways to get results you're looking for. But this might be when I become a buyer of vector DB stock (if I could) when the tide seems to be against them.
While vector DBs might be overkill if you're just searching over a bunch of JSONs, they are super useful when there's more than just text data involved and we're getting ready for a huge influx of video and voice content. GPT-4o is multi-modal and OpenAI is pushing voice. Sora was launched a while ago and we're seeing increasing competition to video and image models. If we're about to be inundated with a ton of voice, video, and text data that we need to efficiently retrieve and serve to users, guess what's going to be the best tool for the job: vector DBs!
No longer will vector DBs be sad that they wouldn't have fulfilled their purpose (pardon my anthropomorphizing); they'll be able to live productive, happy lives helping serve all the data we can possibly generate.
We can have a user ask a question (voice), which directly grabs a video as an answer since we can have everything embedded in the same space and stored in a vector DB. Text might not be the universal interface and vector DBs will have to save us from data overload if that's the case.