skip to content

What does the k constant in reciprocal rank fusion control?

level: middleimportance: must knowfreq 62%

answer

  1. it lives in a denominator
  2. added to the rank before the reciprocal
  3. small values make rank one dominate
  4. large values reward appearing in both lists
  5. the conventional default is sixty

basics

~20 s

Reciprocal rank fusion gives each document 1/(k + rank) from every list it appears in, summed. The constant k controls how sharply top ranks dominate: small k makes rank one overwhelming, large k flattens the curve so agreement across lists matters more.

solid answer

~50 s

Reciprocal rank fusion scores a document as the sum over input lists of `1 / (k + rank)`, where rank is 1-based and lists that omit the document contribute nothing. The constant sits in the denominator, so it damps the head of the curve. With a small k, rank 1 is worth vastly more than rank 5 and a single retriever can dictate the fused order. With a large k, the gap between rank 1 and rank 50 narrows, so a document that both retrievers liked moderately outranks one that a single retriever loved. The value 60 from the original formulation is the usual default and is rarely the thing worth tuning. The deeper property is that RRF reads only ranks, never scores — which is why it needs no normalization between an unbounded BM25 score and a bounded cosine similarity, and why it discards how *confident* either retriever was.

code

python · 7 lines
python
def rrf(ranked_lists, k=60, weights=None):
    scores = {}
    for i, docs in enumerate(ranked_lists):
        w = 1.0 if weights is None else weights[i]
        for rank, doc_id in enumerate(docs, start=1):
            scores[doc_id] = scores.get(doc_id, 0.0) + w / (k + rank)
    return sorted(scores, key=scores.get, reverse=True)

go deeper

for a junior

Memorise the shape of the formula: each list contributes 1 divided by (k plus the document's rank), and the contributions are added up. Know that only positions matter, not the retrievers' scores.

for a middle

Explain what moving k does, with a concrete comparison — one top hit in a single list versus a middling rank in both. Be clear that k damps the head of the curve and is not a depth cutoff.

for a senior

Discuss why rank fusion is the safe default in production (no calibration, no fill values, trivially extensible) and what it costs you: no usable relevance threshold and no way to notice that one leg collapsed.

for a principal

Own when rank fusion stops being appropriate — when the product needs a meaningful confidence number, per-query routing between retrievers, or magnitude-aware blending — and what the team takes on in calibration work by moving to score fusion.

## The formula Reciprocal rank fusion combines several ranked lists into one. For a document `d` appearing at 1-based rank `r_i(d)` in list `i`: ``` RRF(d) = Σ_i 1 / (k + r_i(d)) ``` Documents missing from a list simply contribute no term for that list — there is no penalty, no negative value, and no averaging over the number of lists. The sum is what makes cross-list agreement pay: appearing in two lists means two positive terms. ## What k actually does The constant is added to the rank before the reciprocal, so it is a damping factor on the head of the curve. With **k = 1**, the contributions are 1/2, 1/3, 1/4, 1/5 … Rank 1 is worth 50% more than rank 2 and more than double rank 4. A single retriever's top hit can outweigh strong agreement anywhere below it. With **k = 60**, the contributions are 1/61, 1/62, 1/63 … barely distinguishable. Rank 1 is worth about 3% more than rank 3. Now the count of lists a document appears in dominates almost everything. A worked comparison makes it concrete. Take k = 60. A document at rank 1 in the lexical list and absent from the dense list scores 1/61 ≈ 0.0164. A document at rank 3 in *both* lists scores 1/63 + 1/63 ≈ 0.0317 — nearly double. The mediocre-but-agreed document wins. Set k = 1 instead and the first document scores 1/2 = 0.5 while the second scores 1/4 + 1/4 = 0.5: a tie. Lower k further and the single top hit wins outright. So k is really a knob on **how much you trust one retriever's conviction versus consensus between retrievers**. It is not a depth cutoff — documents at ranks beyond k are not excluded; they just contribute a small, and with large k a nearly uniform, amount. ## Why 60 The value 60 comes from the original formulation of the method and has become the conventional default across implementations. It is a large-ish value, so the default behaviour is consensus-favouring, which is the safe posture when you do not know which retriever to trust on a given query. Treat it as a starting point: it is worth sanity-checking against your own judgment set, but tuning k is usually a second-order gain compared with fixing what each leg retrieves in the first place. ## What RRF gives up RRF is **scale-free**: it never touches the retrievers' scores. This is its whole appeal in hybrid search, where one leg produces an unbounded, corpus-dependent lexical score and the other a bounded similarity. No normalization, no per-query calibration, no fill value for missing documents, and adding a third retriever requires no re-tuning. It is robust in exactly the way that per-query score normalization is not. The cost is that it discards magnitude. A retriever that found one obviously perfect match followed by nine terrible ones is reported to the fusion as "ranks 1 through 10", indistinguishable from a retriever whose ten results are all decent. If the top hit was a verbatim title match and the runner-up was junk, RRF still puts them one small step apart. The consequences show up as: - **No natural relevance threshold.** Fused RRF scores are tiny numbers whose magnitude means nothing absolute, so you cannot cut at a score to decide "no good results". That decision has to be made before fusion, on each leg's own scores. - **Blindness to a collapsed leg.** If one retriever's entire list is irrelevant, its ranks look exactly as authoritative as a good leg's, and it still pushes its candidates into the fused head. ## Weighting and variants A weighted variant multiplies each list's contribution by a per-list weight, letting you say the lexical leg counts more than the dense one on this corpus. This keeps the scale-free property while restoring some control, and is generally the first thing to reach for before touching k. Some implementations also cap the depth of each input list (fuse only the top 50 or top 100 per leg) — that *is* a genuine cutoff, and it is a separate parameter from k, which is a common source of confusion. If you truly need score magnitude — for thresholding, for confidence-weighted blending, for exposing a meaningful relevance number — RRF is the wrong tool and you must normalize scores instead, with all the calibration work that implies. Most production hybrid systems start with RRF precisely because that work is easy to get wrong.

  • Why does reciprocal rank fusion need no score normalization between a lexical and a dense retriever?
    Because it reads only positions. Each list contributes 1/(k + rank), so an unbounded lexical score and a bounded similarity never meet on the same axis. Nothing has to be calibrated, no fill value is needed for documents missing from a list, and adding a third retriever changes no existing parameter. The price is that all magnitude information is thrown away.
  • How do you make one retriever count more than the other under rank fusion?
    Use the weighted variant: multiply each list's 1/(k + rank) terms by a per-list weight before summing. That preserves the scale-free property while letting you encode that, say, the lexical leg is more trustworthy on this corpus. Tune the weights on a judgment set. It is a better first knob than k, which mostly trades single-list conviction against cross-list consensus.
  • Can you use a fused rank-fusion score as a relevance cutoff?
    No. Fused values are small numbers whose magnitude depends on k and on how many lists a document appeared in, not on how relevant it is; the top document of a hopeless query scores much the same as the top document of a perfect one. Any "no good results" decision has to be made on each leg's own scores before fusion, or by a reranking stage after it.

It is like a voting rule where a first-place vote is worth only slightly more than a fifth-place vote: the candidate every judge ranks decently beats the one a single judge adores.

saying these in an interview costs you the question

  • Says k is a cutoff depth that excludes documents ranked below it
  • Thinks reciprocal rank fusion averages instead of summing across lists
  • Believes documents missing from a list receive a penalty
  • Treats the fused score as a calibrated relevance value
  • Claims tuning k is the main lever for hybrid relevance

context

resolving links…