Yes 100% - vectorstores are the bread and butter of enterprise AI solutions.
The solutions I currently recommend to customers are to avoid using LLMs in production altogether for any "reference" type of answers.
The core concept of a RAG system is you hit the vectorstore first, submit that to the LLM, and then send result to the user. But, what becomes obvious is that sending reference data to the LLM is a waste of computing resources.
This is how recommendation systems work - it's actually a "similar" search in a vectorstore.
What is amazing is that data stored in a vectorstore can be anything - images, audio, text, etc... so, "similar" can be anything...
For example - this is how I'd implement a competing product to Shazam - you take the audio, send it to a vectorstore, return the highest confidence result: BOOM.
For the 'No' use case (reference type answers), would you still use a vectorstore or another kind of database?
Yes 100% - vectorstores are the bread and butter of enterprise AI solutions.
The solutions I currently recommend to customers are to avoid using LLMs in production altogether for any "reference" type of answers.
The core concept of a RAG system is you hit the vectorstore first, submit that to the LLM, and then send result to the user. But, what becomes obvious is that sending reference data to the LLM is a waste of computing resources.
This is how recommendation systems work - it's actually a "similar" search in a vectorstore.
What is amazing is that data stored in a vectorstore can be anything - images, audio, text, etc... so, "similar" can be anything...
For example - this is how I'd implement a competing product to Shazam - you take the audio, send it to a vectorstore, return the highest confidence result: BOOM.
No need for an LLM.