Query Decomposition and Augmentation (RAG)

[@ashmit155]

Sorry if i got this wrong but shouldnt query decomposition result in worse results? The whole point behind the transformer model was attention to detail i.e get results and weights of a word according to all the other words/tokens making the model understand (Predict) the next token much more accurately. With query decomposition we are breaking the query and selecting the the result for each query with the highest probability from the database , but shouldnt that result in an output which is not closely related to the other queries? Also would love a hands on rag coding series to understand this more Thanks!

[@MattPayne-b2c]

Hey! I'm covering it a bit more in the next vid for the series, but here's a short description that I was hoping this vid made clearer: With the model focuses in on the semantic similarity between queries and documents, when you have queries that have multiple key topics, they bog each other down when comparing to document chunks. These models look at the entire sentence vector embedding from the query, not individual words. So having multiple topics or ideas means that when comparing to a document that only has one topic you get a lower similarity score. It's no longer a one to one between the two sides, as there is now another topic or idea in the query.