
Agentic Search Architecture: Five Patterns That Survive Production
Agentic retrieval lifts nDCG at 10 by 8.7 points but costs 107.4 seconds per query against 0.67 for one pass. Five patterns keep the gain without the cost.
Tengence Team2026-11-15

Agentic retrieval lifts nDCG at 10 by 8.7 points but costs 107.4 seconds per query against 0.67 for one pass. Five patterns keep the gain without the cost.

Choosing a vector database on a feature checklist is the common mistake. Compare constraints: corpus size, latency budget, hybrid search, existing stack.

Reranker choice is two-stage: a bi-encoder retrieves top-100 and a cross-encoder reranks to NDCG@10 = 74.30. Compare accuracy, latency and business rules.

Vector vs keyword search isn’t either/or. BM25 for exact terms; vectors for synonyms. Hybrid retrieval is the default—BEIR nDCG@10 +5–15% over either alone.