← projects · topic
Query Performance Prediction.
Predicting how well a query will perform — before retrieval, during retrieval, and now: as a selector for the right query variant in a RAG pipeline.
-
%20(1).png)
Can QPP Choose the Right Query Variant?
Pick the winner before paying the RAG cost.
You can over-generate query variants cheaply — but running a full RAG pipeline for each one is expensive. We frame Query Performance Prediction (QPP) as a query selector: given different LLM reformulations of the same information need, can QPP identify which one will produce the best end-to-end RAG answer, without executing the full pipeline for every variant?
+66% end-to-end RAG quality with BM25 (TREC-RAG 2024)+18% end-to-end RAG quality with a dense retrieverEven simple pre-retrieval predictors like IDF work surprisingly well. The bigger lesson: retrieval-optimal ≠ RAG-optimal. The query variant that produces the best-ranked list doesn't necessarily produce the best generated answer — there is a fundamental utility gap between ranking and generation. So don't just optimize retrieval; optimize for answer quality, and let lightweight QPP route to better variants before generation.