find_relevant_tables

Embedding-cosine semantic search over indexed column descriptions.

Layer: Physical schema

Embedding-cosine retrieval over indexed column descriptions. Returns the most relevant tables for a natural-language question, ranked by score. Each hit names the matched column and surfaces its description.

{
  "qualified_name": "public.orders",
  "score": 0.79,
  "best_column": "user_id",
  "best_column_description": "Unique identifier linking each order to the customer who placed it",
  "token_estimate": 49
}

Note on weak matches: A low score (say, < 0.85) is a signal, not an error. The tool returned what semantic search found; the agent should judge whether the score is strong enough to trust. See How retrieval works for why this design.