The team needs to merge BM25 and vector results, but raw scores are not comparable across retrievers. Reciprocal Rank Fusion: combine ranked lists using rank positions, not raw score scales The common trap is averaging BM25 and vector scores without calibration. That makes the larger or more variable score dominate. List ranks BM25: A=1, B=2, C=3. Vector: B=1, C=2, A=3. Use each retriever's rank position; ignore raw score scale. Apply RRF With k=60, A gets 1/(60+1) + 1/(60+3). B gets 1/(60+2) + 1/(60+1). C gets 1/(60+3) + 1/(60+2). Higher ranks contribute slightly more, but agreement across lists matters. Sort fused…
Read the full lesson
Sign up free: one personalized lesson every day, matched to your role and goals.
Already have an account? Sign in