SAGA addresses type-blind grounding in agentic text-to-SPARQL generation. Instead of relying mainly on lexical relevance and observed instances, the training-free framework constrains property exploration with entity types, property domains and ranges, and expected answer types. It maintains a persistent bidirectional type state, removes known-incompatible properties during candidate construction, and presents remaining graph patterns with compact schema annotations. When schema information is incomplete, SAGA permissively supplements it with empirical and trace-local evidence. On nine benchmark settings spanning Wikidata and Freebase, the paper reports the best F1 in all settings, the best exact-match accuracy in eight, and fewer empty-result queries across the reported Wikidata settings.
No heat snapshots are available in the last 24 hours.