Yep! this mostly makes sense with our composability framework, which allows you to define a graph over the data. in the nb linked in the tweet thread here, we define a keyword table index over vector indices.
https://twitter.com/jerryjliu0/status/1635306876358492160?s=20Let's say the overall query is to "compare and contrast between boston and houston sports". If one vector store represents houston and the other represents boston, then when the query hits the vector store (the subindex), we'll want to "decompose" the query into a form that the vector store can answer, e.g. "what is the sports environment of boston like"
