Add reranking when your top-k retrieval (typically k=3-10) frequently includes irrelevant documents in positions 1-3, even though relevant ones exist in positions 4-10. This pattern indicates your initial scoring isn't capturing true relevance. Skip reranking if:
When you do implement reranking, use a cross-encoder model specifically fine-tuned for your domain - generic rerankers add 50-200ms latency but domain-specific ones can boost relevance by 20-30% in production scenarios with complex, multi-faceted queries.