What is the difference between reciprocal rank fusion and reranking?

Reciprocal rank fusion merges two or more ranked result lists into a single unified list; reranking is a separate model-based pass that rescores and reorders one preranked list for better semantic relevance. They are complementary, not mutually exclusive, and are commonly used together in modern retrieval pipelines.

In the RRF formula, the fusion constant is a different concept from the k used by a nearest-neighbour query to request candidates. Follow the documented meaning of each parameter rather than transferring a value between them.

What each stage does

Reciprocal rank fusion (list fusion)

RRF combines already-ranked lists produced by different queries or methods (for example, BM25 text search and one or more vector searches running in parallel). In Azure AI Search, RRF runs whenever there are multiple parallel query executions, such as hybrid search and multi-vector queries, and outputs a single fused list with a unified score for each document (Azure AI Search: Hybrid Search Scoring).

• Inputs: Two or more ranked lists (e.g., BM25 list + vector list[s]).
• Core idea: Assign a reciprocal-rank score per list and sum across lists; the basic scoring uses 1/(rank + k), where k is a constant. Azure documentation notes experiments often use a small k such as 60 for its implementation (Azure AI Search: Hybrid Search Scoring).
• Output: One merged ranking with an overall search score; the magnitude depends on how many lists are fused and their ranks. In Azure responses, RRF produces the @search.score for hybrid results and its upper bound relates to the number of lists being fused (Azure AI Search: Hybrid Search Scoring).
• Notes: In Azure, you can weight contributing queries before fusion to increase or decrease their influence (Azure AI Search: Hybrid Search Scoring).

Reranking (L2 model-based reordering)

Reranking takes a single preranked list (BM25 or already RRF-fused) and applies language understanding models to rescore and reorder the top results for better semantic relevance. In Azure AI Search, semantic ranking is explicitly a secondary pass over an initial BM25 or RRF list, returning a separate @search.rerankerScore and optional captions/answers (Azure AI Search: Semantic Ranking Overview).

• Inputs: The top N items from one ranked list. Azure truncates to the top 50 and uses text-only inputs drawn from configured fields due to model input limits (Azure AI Search: Semantic Ranking Overview).
• Core idea: Deep language models assess semantic relevance given the query and document summaries; they rescore and reorder the candidates.
• Output: A reranked list with @search.rerankerScore in a 0.0 to 4.0 range, plus semantic captions and optional verbatim answers extracted from content (Azure AI Search: Semantic Ranking Overview).
• Limits: This step does not search the whole corpus or add new items; it only reranks the existing top segment (top 50 in Azure’s implementation) and works over text fields. It can rerank pure vector queries if documents include relevant text fields (Azure AI Search: Semantic Ranking Overview).

How they combine in a retrieval pipeline

A common production workflow looks like this:

1. Candidate generation: Run BM25 keyword search and one or more vector searches in parallel.
2. Fusion: Apply RRF to merge the lists into a single, unified ranking and score.
3. Reranking: Feed the top segment of that unified list to a reranker that applies language understanding to rescore and reorder the final output. Azure explicitly documents that semantic ranking occurs after RRF when both are used (Azure AI Search: Hybrid Search Scoring) (Azure AI Search: Semantic Ranking Overview).

For model specifics and tradeoffs, see our separate guide: How do reranker models work?

When comparisons are meaningful (and when they are not)

• Compare like with like: It is meaningful to compare a fused list with and without reranking, holding the candidate pool constant. In Azure, note that RRF’s @search.score and the reranker’s @search.rerankerScore live on different scales and are reported separately (Hybrid Search Scoring) (Semantic Ranking Overview).
• Don’t compare raw scores across methods: BM25, vector similarity, RRF, and reranker scores have different ranges and meanings; direct numeric comparisons across systems are not valid (Hybrid Search Scoring).
• Keep cutoffs consistent: If your reranker operates on top 50, measure quality with the same cutoff when toggling reranking on/off to isolate the effect of the reranker rather than pool size changes (Semantic Ranking Overview).

Practical guidance: when to use which

• Consider RRF when you have multiple ranked inputs to combine (hybrid or multi-vector). It is a deterministic, inexpensive way to reward documents that appear consistently near the top across lists (Hybrid Search Scoring). For context on BM25 as one input, see What is BM25 keyword search?
• Add a reranker when your documents are text-rich and you want better alignment to query intent. Azure notes this step is resource-intensive, works over a truncated top set, and returns semantic captions and optional answers to improve UX (Semantic Ranking Overview).
• Compare the combined pipeline against a simpler baseline: generate with BM25 and vectors, fuse with RRF, then rerank with a semantic model.

Actionable checklist

• Define candidate generators: BM25 fields and one or more vector queries.
• Decide which lists to fuse and any per-list weights before RRF.
• Read the documented fusion method and distinguish its rank-fusion constant from the vector-query candidate count. Do not assume a platform exposes the fusion constant as a setting (Hybrid Search Scoring).
• Choose the reranker’s input fields and top-N cutoff; ensure key text fields are available to the model (Semantic Ranking Overview).
• Evaluate with consistent candidate pools and cutoffs; avoid cross-method score comparisons.
• Measure latency and cost impact of the reranking stage; cache where appropriate.

FAQ

Can I rerank pure vector results? Yes, if documents include semantically rich text fields used by the reranker (Semantic Ranking Overview).

Does reranking find new documents? No. It only reranks the existing top segment (top 50 in Azure’s implementation) rather than searching the full corpus again (Semantic Ranking Overview).

Should I expect universal best RRF settings? No. Follow your platform’s guidance and validate on your data; Azure documentation mentions small k values such as 60 for its implementation (Hybrid Search Scoring).

Related: Semantic search vs hybrid search