When Reranking Becomes a System Boundary

Author: Ravindra Harige Source: https://www.searchplex.net/blog/when-reranking-becomes-a-system-boundary Published: 2026-05-25

Summary

A reranker only sees what retrieval allows to survive. Once the reranker becomes the dominant driver of final ordering quality, the system has quietly transferred the relevance problem downstream — and no individual team’s dashboard shows the full picture. This piece names that boundary and explains why most teams cross it before they recognize it.

Core Argument

Retrieval Defines Eligibility, Reranking Defines Order

If a document is not retrieved, it never participates in ranking. No downstream stage can reintroduce it. This creates a hard constraint:

  • Retrieval controls what enters the system (recall)
  • Reranking controls how that subset is ordered

A strong reranker can produce good top-k results even when retrieval is incomplete — but the system may still miss relevant documents entirely.

The External Reranker Structural Tradeoff

Retrieval computes rich query-time signals: term matches, field-level contributions, proximity, BM25 components — how the document matched the query, not just what it contains. External rerankers typically receive only a reduced representation: document text, metadata, embeddings, and a limited feature set.

more candidates → less per-candidate retrieval context
more context → fewer candidates

Ranking as a Projection

Retrieval-time signals are computed once, against the full index. Reranking does not redo that computation — it works with what survives into the candidate set. This lossy compression is structural, not a failure of implementation.

Every gain the reranker makes operates within the bounds of what that projection preserved.

When Reranking Becomes Compensatory

The system crosses a boundary when:

  • Performance improves by widening the rerank window, not by improving retrieval
  • Window size is load-bearing, not just a latency knob
  • Retrieval is the constraint, not the quality driver

A query like red waterproof hiking shoes size 11 illustrates the ceiling: lexical retrieval preserves attribute constraints but misses semantic variants; semantic retrieval captures relevant footwear but loses attribute precision. The reranker cannot recover what was never retrieved.

Evaluation Splits Across Stages

StageMetricWhat It Measures
RetrievalRecall@KWhether relevant docs appear in candidate set
RerankingNDCG, MRRHow well the candidate set is ordered

These are partially independent. NDCG can improve while recall is weak, and retrieval improvements may not immediately improve top-k ordering.

Asymmetric visibility failure: NDCG improves → team declares success. User-visible relevance plateaus. Neither team’s dashboard shows the full picture.

Organizational Ownership Split

StageOwnerOptimizes For
RetrievalEngineeringIndexing correctness, query latency, recall, reliability
RerankingML / Data ScienceModel quality, feature engineering, offline ranking metrics

A retrieval change shifts the input distribution the reranker was tuned against. A reranker improvement masks retrieval weaknesses, removing pressure to fix them. Each system looks correct under its own metrics, with its own assumptions about what the other stage provides.

Local correctness does not imply global correctness. That gap closes only when someone is accountable for the space between them.

Key Insight

When reranking is asked to compensate for weak retrieval rather than refine a good candidate set, the system has transferred the relevance problem to a stage operating on a lossy projection of retrieval-time signals, optimized by a team that does not own the stage it depends on most.