There is no universally best way to fuse lexical and vector retrieval. Use reciprocal rank fusion (RRF) as a practical starting point when component scores are not comparable or you lack reliable relevance labels; test weighted score fusion when score margins carry useful signal and you can normalize and tune them; consider learned fusion when you have representative judgments and can maintain a rigorous training and evaluation loop. The winner depends on the retrievers, corpus, query mix, and ranking cutoff.
What the three fusion methods do
Reciprocal rank fusion (RRF)
RRF combines the positions of a document in multiple ranked result lists. A common formula is score(d) = Σ 1 / (k + rank(d)), where rank(d) is the document’s position in a list and k controls how much position affects its contribution. A document appearing near the top of several lists can accumulate a strong fused score.
As an Amazon Associate I earn from qualifying purchases.
RRF uses ranks, not the component scores’ magnitudes. It therefore avoids directly comparing, for example, unbounded BM25 scores with bounded vector similarities. But it discards score margins: a document that narrowly beats another within one retriever gets the same rank-based contribution as one that wins by a large margin. OpenSearch describes RRF as a reasonable starting point before score distributions have been measured or calibrated (OpenSearch documentation; publication year not stated).
Free tools Windows power users keep installed
One-click scans. No signup required.
RRF is a fusion step, not a relevance model that repairs weak retrieval. The candidates and candidate depth supplied by each retriever still matter: fusion cannot surface a relevant document absent from all input lists. In Azure AI Search, Microsoft describes semantic ranking as a separate reranking stage that can follow RRF, rather than as part of RRF itself (Microsoft Azure AI Search documentation).
#1 Best Overall
Weighted score fusion
Weighted score fusion combines component scores—often after normalization—using a weighted sum or convex combination. Unlike RRF, it can preserve information about the size of score gaps. The trade-off is that score scales, normalization, and weights all need deliberate treatment. Normalizing scores does not by itself show that a particular weighting works; evaluate both on the target data. OpenSearch documents score-based normalization and weighted combinations as alternatives to rank-based fusion (OpenSearch documentation).
Bruch, Gai, and Ingber’s 2022 study compared reciprocal rank fusion with convex combination (CC) for lexical and semantic retrieval. In their in-domain and out-of-domain experiments, CC outperformed RRF, and they found its parameter could be tuned with a small set of labeled queries. Their result supports testing a tuned weighted blend; it does not establish that weighted fusion wins on every production corpus.
Learned fusion
“Learned fusion” describes a family of methods, not one fixed algorithm. A system might fit fusion weights from query-document relevance judgments, feed component scores into a learning-to-rank model, or learn a query-dependent rule. These approaches can capture more than one global blend, but their quality depends on representative training data and held-out evaluation. They also add work to train, validate, monitor, and maintain the model.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →The available educational material describes learned ranker-score weighting and using ranker scores as learning-to-rank features; Bruch, Gai, and Ingber provide evidence for the narrower case of tuning a convex combination from labeled queries. Neither establishes that a richer learned fusion method universally outperforms RRF or a tuned global blend.
Which method should you test first?
| Situation | Start by testing | Why |
|---|---|---|
| Score scales are incompatible, relevance judgments are unavailable or scarce, or you need a low-tuning baseline | RRF | It combines ranked positions without requiring directly comparable component scores (OpenSearch documentation). |
| Score margins appear informative, and you can normalize scores and validate weights | Weighted score fusion | It retains score-margin information and lets you tune the balance (OpenSearch documentation). |
| You have representative relevance labels and can support training and held-out evaluation | Learned fusion | It can fit weights or a more expressive ranking rule to observed relevance (Bruch, Gai, and Ingber, 2022; educational lecture material). |
| You do not know which retrieval signal helps which queries | Compare all three on held-out query slices | Which ranking works best is an empirical question for your corpus and query mix. |
Use this as a test plan, not an algorithmic law. If relevant items are missing from the retrievers’ candidate lists, changing the fusion rule alone will not retrieve them.
What published comparisons show—and what they do not
Published results illustrate why a local comparison matters: outcomes vary by task and setup. OpenSearch reports that, across six BEIR datasets in its cited comparison, RRF had an average NDCG@10 3.86% lower than its score-based hybrid pipeline; the documentation says latency and coordinator-node CPU utilization were comparable. This is a result of that cited benchmark, not a forecast for another implementation or workload (OpenSearch documentation; publication year not stated).
MTEB’s documentation gives task-specific NDCG@10 values for BM25, a dense model, RRF, DBSF, and RSF. The documented hybrid models use equal weights. The different outcomes across these three tasks are evidence against treating one task’s result as a universal ordering:
Recommended Free Tools
| MTEB task | BM25 | Dense | RRF | DBSF | RSF |
|---|---|---|---|---|---|
| NanoSciFactRetrieval | 0.710 | 0.725 | 0.754 | 0.538 | 0.767 |
| NanoNFCorpusRetrieval | 0.325 | 0.288 | 0.329 | 0.338 | 0.359 |
| NanoSCIDOCSRetrieval | 0.335 | 0.344 | 0.369 | 0.344 | 0.372 |
These are the values in MTEB’s documented example; the documentation page does not state a publication year. They are task-specific NDCG@10 results, not a guarantee about other corpora or differently configured systems (Massive Text Embedding Benchmark documentation).
Best Value
In the 2022 convex-combination study, the authors report that its fusion parameter converged with less than 5% of the training data for the datasets they studied. That qualification applies to those datasets and that parameter-tuning setup; it is not a general sample-size requirement for learned fusion. The original RRF paper by Cormack, Clarke, and Buettcher (SIGIR 2009) reports favorable results against individual systems and standard Condorcet Fuse in its experiments. That abstract is not a modern head-to-head verdict against weighted or learned hybrid fusion.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to compare fusion methods on your system
- Fix the retrieval setup. Hold the lexical and dense retrievers, corpus, candidate depths, and judged queries constant. Otherwise, you cannot attribute a change in results to the fusion method.
- Separate tuning from evaluation. Split representative judged queries into tuning and held-out sets. Tune weights or learned models only on the tuning portion, then compare all candidates on queries they did not use for tuning.
- Choose a metric and cutoff that match the product. Measure ranking quality at the depth where relevance affects the application. NDCG@10 is one example used in the MTEB documentation, not a mandatory metric for every system.
- Inspect query slices, not only the aggregate. Report results for query forms such as exact names or identifiers, short keyword searches, and longer natural-language requests. A single aggregate can hide cases where one retrieval signal helps and another hurts.
- Measure operational trade-offs. Include serving cost, latency, score stability, and the frequency of recalibration or retraining in the comparison. OpenSearch’s cited BEIR result found comparable latency and coordinator CPU between the pipelines it compared; measure these costs in your own implementation and workload.
- Re-evaluate after meaningful changes. A new corpus, query mix, or component retriever can change which signals are useful. Do not transplant a published weight or RRF parameter without validation.
When each method is most defensible
Choose RRF for a robust first baseline
RRF is a sensible first experiment when score scales are incomparable, score outliers make direct combination risky, labels are limited, or you need a low-tuning baseline. Its simplicity does not mean its behavior is parameter-free: rank contributions depend on the chosen setup, and Bruch, Gai, and Ingber found RRF sensitive to its parameters in their experiments.
Choose weighted fusion when margins have signal
Test a weighted blend when score differences carry information that ranking positions would lose and you have a stable normalization strategy. Use representative relevance judgments to tune and validate the weights, and check performance on held-out queries rather than treating normalization as proof that the blend is calibrated.
Choose learned fusion when data and operations support it
Consider learned fusion when your judgments cover the query types and relevance outcomes that matter, when query types may benefit from different treatment, and when the team can maintain an evaluation and retraining process. Compare it with both RRF and a tuned global weighted blend; added model complexity is useful only if held-out quality and operational results justify it.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
