Expand description
Several stores, one query.
A developer’s work is not one repository, so a question about “the code” is
rarely a question about one store. A Fleet opens the stores it is given,
asks each of them the same question, and merges the answers into one ranked
list where every hit says which store it came from.
Two properties of the store make this cheap. Chunk ids never leave a store — a hit is resolved to text inside the store that produced it — so no id from one store can be looked up in another. And a search’s score is a reciprocal-rank sum, not a distance, so scores produced by two different stores with two different models are still the same unit.
Structs§
Traits§
- Labelled
- A graph row that can say which store it came from.