Source: Practitioner pattern across e-commerce search (Bass Pro Shops, Best Buy, Coveo deployments); enterprise search platforms; documented in Grainger's AI-Powered Search
Classification — Pattern for selecting retrieval architecture based on query intent classification.
Route each incoming query to the most appropriate retrieval pipeline based on classified intent (navigational, informational, conversational, transactional), producing better per-query results than uniformly applying one architecture.
Queries vary substantially in what retrieval they need. A navigational query ("nike air max") wants exact-match retrieval that surfaces the specific known item. An informational query ("red running shoes") wants hybrid retrieval with facets for refinement. A conversational query ("which shoes are good for marathons?") wants RAG-style retrieval with reranking and possibly answer synthesis. A transactional query ("buy nike air max 270 size 10") wants filter-heavy retrieval. Applying one uniform architecture to all of these compromises every category.
Intent signal source. Intent classification (covered in the future Query Understanding Catalog) produces the routing signal. The classifier may be rule-based (length, structure, keyword patterns), classifier-based (a small LLM or trained model that outputs intent), or learned (a model trained on query/intent labels). Production deployments often combine: rule-based for clear cases, classifier-based for ambiguous cases.
Route definitions. Each intent maps to a retrieval architecture. Navigational → lexical-heavy with aggressive exact-match boosting. Informational → hybrid retrieval with facet generation. Conversational → dense retrieval with cross-encoder reranking. Transactional → filter-applied retrieval with structured constraint handling. The mappings are implemented as separate retrieval pipelines that the router selects among.
Confidence-based fallback. The router's intent classification has confidence scores. High-confidence classifications route directly to the designated pipeline. Low-confidence classifications can: run multiple pipelines in parallel and fuse the results (defensive routing); route to a default "general" pipeline (conservative routing); or use the broadest architecture (recall-preserving routing). The choice affects quality and cost.
Per-route tuning. Each retrieval pipeline can be tuned independently. The navigational pipeline's BM25 parameters, exact-match boosts, and facet handling are tuned for known-item queries. The informational pipeline's hybrid weights are tuned for discovery queries. Per-route tuning is a substantial advantage over uniform architectures, where any tuning is a compromise across query types.
Operational complexity. Routing adds infrastructure: the router itself, multiple retrieval pipelines, evaluation per route, monitoring and alerting per route, deployment management for multiple pipelines. The complexity is justified when the workload's heterogeneity is significant; deployments with mostly one query type often don't justify routing.
Monitoring and analytics. Per-route metrics (latency, recall, click-through, conversion) reveal whether routing is working. Misrouted queries appear as anomalies: high-CTR queries in the wrong pipeline; low-CTR queries that should have been routed differently. The discipline of monitoring routing decisions is itself a search operations pattern; the future Search Operations Catalog covers it.
E-commerce search with mixed query types (navigational, informational, transactional). Enterprise search with diverse content types and use cases. Consumer search platforms where the query distribution spans many intents. Cases where evaluation data shows that one-size-fits-all retrieval produces visibly worse results for specific query types.
Alternatives — single-architecture deployment for narrow workloads with uniform query types. Federated multi-index retrieval (next entry) for cases where content heterogeneity matters more than query heterogeneity. Per-query hybrid weighting (Section C) for lighter customization without full routing infrastructure.
- Trey Grainger, AI-Powered Search, chapters on query understanding and routing
- Coveo machine learning query intent documentation
- Practitioner case studies (Bass Pro Shops, Best Buy, etc.) on e-commerce search routing