Multi-Rep & ColBERT Retrieval Fashions for RAGs

Listed below are my favourite methods — one is quicker, the opposite is extra correct.

I construct RAG apps; it’s enjoyable!

However the apps I construct don’t do effectively in manufacturing. They’re promising prototypes, however they by no means go reside!

The wrongdoer is sort of at all times the retrieval. Come on, that is the guts of RAGs. What are we supposed to construct with out this?

That is till I index paperwork for sooner or higher retrieval.

Indexing helps us engineer options that retrieve knowledge sooner. It considerably reduces latency, bettering the general app expertise. We use indexing in nearly each app we construct. It has nothing to do with LLMs or RAGs.

Nearly all of the databases ship with indexing help. As an illustration, Postgres can do B-Tree, GiST, SP-GiST, BRIN, GIN, and Hash forms of indexing. That’s an inventory lengthy sufficient to go to a separate future publish.

On this publish, I’ll talk about the favored indexing methods I regularly use for higher doc retrieval. These methods are, nonetheless, particular to RAG apps. You’ll see why in a second.

My two go-to indexing methods are multi-representation and ColBERT. These aren’t the one strategies now we have. And it’s…