Source: Production methodology at major personalized search companies; emerging methodology on contextual NLP
Classification — Pattern for using user, session, and environmental context to disambiguate query understanding outputs.
Adjust query understanding outputs — intent classification, entity linking, synonym expansion — based on context signals about the user, session, and environment, producing per-user-per-context understanding that uniform understanding misses.
Query understanding without context produces a single interpretation that may not fit the user. "Jaguar" classified as informational without context could be about the car or the animal; the right answer depends on the user. "Mens" as a standalone query is ambiguous (gender? section? brand?); in session context after "running shoes" it's clearly a refinement. Context-aware understanding handles these cases; context-naive understanding doesn't.
Context signal sources. User-level: prior queries (does this user typically search for cars or animals?); prior purchases or interactions (a cars enthusiast has bought car-related products); demographic signals where legitimately available. Session-level: queries earlier in the session; filters or facets applied; categories of items viewed. Environmental: locale (different countries have different brand prevalence); time (some queries shift meaning by time of day); device (mobile users may shorten queries differently).
Context-aware intent classification. The intent classifier (Section D) takes the query plus context as input. The same query "jaguar" might be classified as transactional with high confidence in a cars-shopping session, and as informational in a wildlife-research session. Production implementations: include context features in the classifier's feature vector (rule-based and ML-based methods); include context in the LLM prompt for LLM-based classification.
Context-aware entity linking. When the query references an entity that could resolve to multiple IDs, context picks the right one. "Jaguar" in a car-shopping context links to brand_id_jaguar_motors; in a nature-research context links to species_id_jaguar. The disambiguation uses context features alongside the query span. Production implementations typically rank candidate links by combining base linking score (how well the surface form matches the ID) with context-adjusted scores.
Context-aware synonym selection. Some synonyms apply only in specific contexts. "Trainer" might mean running shoes in athletic context, but personal trainer in fitness-services context, but training data in ML context. Context-aware synonym selection chooses the appropriate expansion based on context. Implementation: maintain context tags on synonym rules; only apply rules whose tags match the query context.
Session-based query interpretation. Sequential queries in a session often build on each other. "Running shoes" → "mens" → "red" → "size 10" — each subsequent query refines the previous. Production patterns: maintain session context across queries; merge subsequent queries with the session context for retrieval (each query effectively appends to the context). The user experiences progressive narrowing; the system handles each query in light of what came before.
Cold-start and degradation. New users without history, anonymous sessions, no available context: context-aware understanding needs to degrade gracefully. Fall back to context-naive understanding. Use weak available signals (locale, device) cautiously. Build up session context as queries accumulate.
Privacy considerations. Context use can feel intrusive if users perceive that the system is making assumptions about them. Best practice: keep context use visible when possible ("showing results for cars" when the system inferred cars from session context); allow easy override ("search for jaguar in nature topics instead"); avoid context use that the user couldn't consent to (using inferred demographic signals to disambiguate). The discipline overlaps with broader personalization ethics from Volume 4 Section E.
Production search where queries are commonly ambiguous and context signals are available. E-commerce with logged-in users where session and purchase history disambiguate. Consumer search where session context strongly informs intent. Multi-domain search where domain context disambiguates entity references.
Alternatives — context-naive understanding for narrow domains where queries are unambiguous. Light context use (locale only) for privacy-sensitive deployments. The right level of context use depends on the workload and the privacy/ethics considerations specific to the deployment.
- Production methodology writings on contextual query understanding
- Coveo personalization and context documentation
- Algolia personalization and rules documentation