Compute the weighted hybrid fusion score for each candidate and give the final ranking
A hybrid retriever combines normalized dense and sparse scores as fused = alpha*dense + (1-alpha)*sparse, with alpha = 0.6. Three candidates have (dense, sparse) scores: A = (0.9, 0.2), B = (0.5, 0.9), C = (0.7, 0.4). Give each fused score and the final ranking, best first.
Blend each candidate as 0.6·dense + 0.4·sparse: A = 0.62, B = 0.66, C = 0.58, so the ranking is B > A > C — B wins despite A's higher dense score because B's strong sparse signal carries enough weight.
Imagine grading three students on two tests, where the first test counts for 60 percent and the second for 40 percent. You don't just look at who aced one test — you blend the two with those weights to get a final grade. Student A crushed test one but bombed test two; student B was so-so on test one but excellent on test two; student C was middling on both. When you mix the scores with the 60/40 weighting, B comes out on top even though A had the single highest score on any test, because B's strong second-test result counts for enough. Hybrid retrieval scoring works the same way: each document gets a blended grade from two search methods, and you rank by the blend, not by either score alone.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Hybrid retrieval is the default in serious RAG systems because dense and sparse search fail in opposite ways: embeddings capture meaning but miss exact terms like error codes or product SKUs, while keyword search nails exact terms but misses paraphrase. Combining them recovers both, and weighted score fusion is the simplest way to do it. This question looks like arithmetic, but the interviewer is using the numbers to check whether you understand what the blend actually does.
The specific scenario is deliberately rigged so the naive answer is wrong. Candidate A has the single highest score in the whole problem — 0.9 on dense — and a candidate who pattern-matches on "highest number wins" will rank A first. But A's sparse score is weak, and once the fusion weights are applied, A loses to the more balanced candidate B. The whole point of the question is that you must compute the blend, not eyeball the inputs.
This deep dive does the arithmetic carefully, draws out why balance beats a spike at this alpha, names the two assumptions the formula quietly relies on, and shows how to reason about alpha as a continuous knob — including solving for the exact crossover where the ranking flips.
The fusion formula and the per-candidate arithmetic
Weighted score fusion combines two normalized retriever scores into one. The formula is a convex combination controlled by a single mixing weight alpha:
With alpha = 0.6, the dense score is weighted 0.6 and the sparse score 0.4. Because the two weights sum to 1, the fused score stays on the same 0-to-1 scale as the inputs, which keeps it interpretable.
Now apply it to each candidate, one multiply and add per term:
For A with (dense, sparse) = (0.9, 0.2): 0.6×0.9 = 0.54, and 0.4×0.2 = 0.08, summing to 0.62.
For B with (0.5, 0.9): 0.6×0.5 = 0.30, and 0.4×0.9 = 0.36, summing to 0.66.
For C with (0.7, 0.4): 0.6×0.7 = 0.42, and 0.4×0.4 = 0.16, summing to 0.58.
Sorting the fused scores best first gives B (0.66), then A (0.62), then C (0.58): the ranking is B > A > C. The most common arithmetic slip here is mis-multiplying 0.6×0.9 as 0.56 instead of 0.54, or applying the 0.6 weight to sparse by mistake — both flip the final order, so it is worth writing each product out rather than doing it in your head.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- Weaviate and Elasticsearch hybrid search expose an alpha (or weight) parameter that blends dense and BM25 scores exactly this way
- Pinecone's hybrid search uses a convex weighting of dense and sparse vectors with a tunable alpha for the same purpose
What an interviewer would ask next. Try answering before peeking at the approach.
QAt what value of alpha do A and B swap rank?
Set the two fused scores equal: 0.9·alpha + 0.2·(1−alpha) = 0.5·alpha + 0.9·(1−alpha). Solve for alpha. You get alpha ≈ 0.636, so for alpha above that A overtakes B and below it B leads. Reporting the crossover, not just the ranking at 0.6, shows you understand alpha as a continuous knob.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Ranking A first because it has the highest dense score (0.9), without computing the blend — at alpha = 0.6 the fused score is 0.62, which B's 0.66 beats.
60 second bullets to scan on the way to the call.
The weighted fusion formula and which weight applies to dense versus sparse
How to compute each candidate's fused score from its score pair
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.