How Onibex ASK Uses a Semantic Layer to Fix Text-to-SQL Failures on SAP Data

Onibex has built an AI agent called ASK (Agentic Semantic Knowledge) that converts natural language questions into accurate SQL queries over SAP enterprise data. The core challenge it addresses is that standard Text-to-SQL systems often hallucinate table and column names when faced with SAP's thousands of cryptic, performance-optimized tables. Instead of letting the language model invent database structure, ASK constrains it to a curated semantic layer made up of YAML-based 'Data Products' that map business terms directly to real SAP columns and join conditions. The system uses a three-tier Bronze-Silver-Gold medallion architecture, where Gold entities are pre-joined and pre-aggregated for fast single-scan queries, while Silver entities handle more complex joins when needed. This design makes generated SQL reproducible and auditable, with the agent asking for clarification rather than guessing when a requested term is absent from the semantic layer.