Watch the silent 512-token truncation when bolting a cross-encoder reranker onto RAG

Starmorph·high signal

A two-stage hybrid-retrieve-then-rerank pipeline (dense + sparse fused with Reciprocal Rank Fusion, then rerank ~20 candidates) lifts recall 0.72→0.91 and precision 0.68→0.87, but most cross-encoders truncate silently at 512 tokens — so longer chunks get scored on a fragment and you never see the error. For longer content use Cohere Rerank 3.5 (4,096-token limit, ~$1/1K requests) or Jina Reranker v3 instead. The truncation gotcha is the non-obvious failure mode that quietly caps reranker quality in production.