From SQL to Semantic Views: 5-Layer Knowledge Engineering for AI Agents

Historical SQL is a gold mine, but feeding it directly to an Agent turns it into noise: duplicates, expired logic, hard‑coded values, and inconsistent calibers may retrieve logic abandoned years ago. A more effective path is a five‑layer refinement process.

Layer 1 – Collection & Clustering: Gather high‑frequency queries and core report SQL, then cluster them by business domain.

Layer 2 – Extraction: From the clustered SQL, extract metric expressions, filter conditions, join relationships, and time calibers.

Layer 3 – Semantic View Organization: Organize the extracted elements into semantic views that explicitly define metrics, dimensions, entities, relationships, and rules. A semantic view is not a physical table; it is an intermediate layer that lets an Agent understand how business objects are calculated, joined, and filtered.

Layer 4 – Physical Mapping & Validation: Map the semantic view back to physical fields. Validate the mapping through historical result comparison, sampling checks, and boundary‑scenario testing.

Layer 5 – Governance: Establish versioning, ownership, and change records to enable continuous maintenance. The article emphasizes that unvalidated knowledge is merely another form of guesswork.

This five‑layer approach is a recurring challenge that the Yunqi product team tackles in business knowledge construction for AI agents.