Multi-dataset Topic best practices for Amazon Quick Chat

This article provides best practices for configuring Amazon QuickSight multi-dataset Topics with AI-generated SQL, enabling natural-language analytics across multiple datasets without pre-defined join graphs.

• Chat-driven SQL generation infers joins at query time based on semantic metadata, supporting outer joins, unions, and subqueries unlike defined-relationship Topics
• The Semantic Guidance Stack comprises seven layers: dataset instructions, topic instructions, synonyms, field descriptions, semantic types, column exclusions, and calculated fields
• Write dataset-level instructions as data dictionaries specifying grain, primary keys, foreign key hints, business rules, and aggregation semantics
• Design topic-level instructions for cross-dataset logic, disambiguation rules, default join behavior, and multi-fact resolution
• Create comprehensive synonyms mapping business vocabulary to technical fields across executive, analyst, domain, and abbreviation tiers
• Enrich field descriptions with definition, unit, nullability, valid ranges, and aggregation behavior for AI consumption
• Guide join behavior through implicit hints, grain alignment, union instructions, subquery patterns, and conditional join logic
• Handle complex patterns: outer joins preserve unmatched records, bridge tables enable many-to-many relationships, self-joins traverse hierarchies, role-playing dimensions serve multiple contexts, and cross-grain comparisons require rollup instructions
• Reduce noise by excluding surrogate keys, ETL metadata, deprecated fields, and low-value high-cardinality columns
• Test semantic models iteratively using tiered question banks from basic single-dataset queries to complex multi-dataset patterns, validating join correctness, aggregation, dataset selection, filters, and NULL handling
• Choose defined relationships for governed environments, semantic-only Chat for exploratory analytics, or hybrid approaches combining both strategies

Precise semantic metadata enables AI-driven SQL generation to deliver accurate results across complex multi-dataset scenarios without structural constraints.