점수 병합을 위한 상호 순위 융합

Dense retrievers produce cosine similarity scores between -1 and 1, while BM25 produces unbounded positive scores that depend on corpus size and term frequencies. You cannot add these numbers directly — a BM25 score of 3.7 and a cosine similarity of 0.85 mean completely different things. Simple score normalization (dividing by max score) is brittle because outlier documents distort the scale. We need a method that is agnostic to absolute score values.

Reciprocal Rank Fusion (RRF) sidesteps the score incompatibility problem by converting each retriever's results into ranked positions and fusing those ranks rather than raw scores. A document ranked first gets a high RRF contribution, a document ranked tenth gets a much lower one, and the final score is the sum of RRF contributions across all retrievers. The formula is: RRF(d) = sum(1 / (k + rank_i(d))) where k is a smoothing constant (typically 60).

The implementation is surprisingly simple. For each retriever, iterate over its ranked result list and accumulate RRF scores into a dictionary keyed by document ID. Documents that appear in multiple retrievers accumulate contributions from each. Finally, sort by total RRF score descending to produce the merged ranking.

In practice, you run both retrievers, collect their ranked document ID lists, pass them to RRF, and look up the top-K document IDs in your document store. Notice that no score normalization is required — only the rank ordering of each retriever's results matters. This makes RRF extremely robust to distribution shifts and corpus size changes.

Notice the pattern of retrieving 20 candidates from each retriever and then taking the top 5 after fusion. This over-retrieval strategy is important because a document that appears at rank 10 in BM25 but rank 1 in dense should bubble up in the fused list. If you only retrieve 5 from each, you miss such documents. A common practice is to retrieve top_k * 4 from each retriever before fusing and then return top_k final results.

The constant k in the RRF formula controls how much weight is given to top-ranked versus lower-ranked documents. A small k amplifies the difference between ranks (rank 1 is much better than rank 2), while a large k flattens the distribution (all ranks matter roughly equally). The original paper used k=60 and empirical studies show values between 30 and 100 all perform similarly. k=60 is a robust default.

RRF handles missing documents gracefully: a document that appears in only one retriever still accumulates that retriever's RRF contribution. It simply does not receive a second contribution. This means documents that both retrievers agree on will naturally score higher than documents found by only one. This consensus signal is exactly what you want in a hybrid system.

RRF scales naturally to three or more retrievers. You might combine BM25, a dense embedding retriever, and a sparse learned retriever (like SPLADE) that explicitly optimizes for keyword matching in a learned sparse space. Each additional retriever adds another RRF contribution per document, and documents that rank highly across all retrievers win by a clear margin.

Standard RRF treats all retrievers equally, but in practice one retriever may be more reliable than another for your specific domain. Weighted RRF multiplies each retriever's contribution by a weight before summing. A weight of 1.5 on the dense retriever and 0.5 on BM25 emphasizes semantic matching, while equal weights of 1.0 treat them symmetrically. Tune weights on a validation set.

Alternatives to RRF include min-max normalization (scale each retriever's scores to [0,1]) and z-score normalization. Both are sensitive to outlier documents that inflate the maximum score and collapse other scores toward zero. RRF is robust to outliers because it only uses rank positions, not raw scores. Empirical benchmarks consistently show RRF outperforms normalization-based fusion approaches.

LangChain's EnsembleRetriever implements RRF fusion under the hood. It accepts a list of retrievers and optional weights, runs them in parallel, and returns RRF-merged results. This lets you integrate hybrid search into LCEL chains with minimal boilerplate and swap out individual retrievers without changing the fusion logic.

Test your understanding of reciprocal rank fusion from this lesson.

In this lesson you learned: RRF merges ranked lists without normalizing incompatible scores, the formula 1/(k + rank) gives each document a contribution from each retriever, and documents ranked highly by multiple retrievers earn the highest fused scores. The smoothing constant k=60 is a robust default. LangChain's EnsembleRetriever implements RRF natively. Next up we configure hybrid search directly in Pinecone and pgvector.