A Deep Dive into Vector Databases – Through a Recommendation Engine Example

The problem statement

In this current landscape of LLM powered applications where response generation is mostly domain centric and has to be trusted, vector database plays a critical role. Vector database allows for storing knowledge repositories and grounding the context for LLMs to generate correct answers. In this blog I am not going to discuss the vector database selection process of how to design one ( there are enough content out there and I am planning to write a dedicated blog around that soon). However, in this blog I want to discuss what should be the characteristic of a holistic vector database and what kind of search it should support at the minimum.

In my opinion, one true way to analyze that is through a real world scenario where we might really need to take the vector database for a test ride and I have just the use case !

A recommendation engine ( powered by LLMs) need to show the depth and breadth of multiple kinds of searches. For example we need to do a simple text search, a semantic search, filter based search and personalized search.

So, let’s evaluate the “art of possible” for vector search and see what we might need from a true vector database.

The solution

A holistic recommendation engine should support multiple kinds of search as below ( not exhaustive of course )

• Simple text based search