Skip to main content
Layer: Semantic layer
Embedding-cosine retrieval restricted to entities. Reuses the column-level embedding index (no second model) but ranks per entity by taking the MAX cosine across the columns of each entity’s bound table. Returns domain-named hits so the agent stays in business terms. 
{
  "name": "customer",
  "score": 0.84,
  "qualified_table": "public.users",
  "best_column": "email",
  "best_column_description": "Primary contact email used for order confirmations and password reset",
  "token_estimate": 58
}
Empty envelope routes the agent differently depending on whether the semantic layer is bare or just unmatched:
  • No entities curated yet → follow_up_hints: ["find_relevant_tables"] (skip list_entities, it would also be empty).
  • Entities exist but none match → follow_up_hints: ["list_entities", "find_relevant_tables"].
Use find_relevant_tables instead when no entities are curated.