At depth 30 the two sets line up exactly, and that is what makes the shortcut tempting. One question was wrong, and exactly one question's answer passage fell outside the candidate set [1][4]. A passage the reranker never sees cannot be promoted, so the out-of-candidates set is always a subset of the miss set, and when both have size one they are the same set. The author of the dev.to writeup drew the conclusion that follows: the reranker promotes whatever it is handed, so depth is the bottleneck [16]. That inference is valid only while the two sizes match. Depth 100 empties the out-of-candidates set and leaves the miss set at one, and the conclusion dies with it [6].
Now the cost side. A reranker here is a cross-encoder, so handing it 100 candidates means 100 query-passage relevance computations; the bill is linear in candidate count [8]. The measured numbers are worse than linear. Per question, depth 30 took 6.8 seconds and depth 100 took 37.7 [7]. Scaling the smaller figure by 100/30 predicts 22.7 seconds, so the run came in about 1.66 times over that prediction [3]. Fit both points to a fixed overhead plus a per-candidate cost and you get 0.441 seconds per candidate and an overhead of minus 6.4 seconds [5], which no profiler will sell you. Two data points is not a model, but on this CPU ONNX path the marginal candidate got dearer rather than cheaper. Across 25 questions, depth 100 spent roughly 12.9 extra minutes of wall clock [2] for no additional correct answers [6]. The plan's cost line said "zero cost, zero VRAM"; the VRAM half holds [9].
The benchmark that justified the sweep is a public set of 3,493 traditional-Chinese questions where a weak embedding plus a reranker reaches 98%+ at k=100, beating every single-stage retriever [13]. For that number to transfer, your reranker has to be willing to pull a candidate the embedding ranked 83rd into the top few. The one in this stack is bge-reranker-base, a 2023 base-tier model, and it saw that passage and left it outside the top 5 [6][11]. The plan also carried a SIGIR 2025 warning about over-reranking degrading results [14]; the failure mode that actually showed up was not degradation but flat accuracy at 5.5 times the latency [7].
The named fix does not install. The to-do said move to bge-reranker-v2-m3, billed as +1.3GB, same architecture, same API, no prompt changes [10]. fastembed's cross-encoder list has six models and v2-m3 is not among them; only two are multilingual, the incumbent and jinaai/jina-reranker-v2-base-multilingual at 1.11 GB [11]. Getting v2-m3 means moving the whole service off ONNX onto PyTorch [12], which is not a config change.
One caution on resolution. The measured table is 25 questions, where a single question is 4 percentage points [18], while the sweep plan describes 29 questions per configuration [2]. At that size the ceiling column is a bound, not an accuracy estimate. What the probe buys is knowing which of the two conditions you are failing: seeing the passage at all, or getting it ranked [17].
