Ask the Catalog

Natural-language search across Unity Catalog, Azure Purview, and ADF metadata — answers cite real tables, ranked by actual query usage.

How this works

LLM over embedded catalog metadata + query history. Ranking blends semantic match (60%) with usage-based relevance (40%) — tables that people actually query rank higher than tables that just look similar.

Tables indexed

1,284

across 3 catalogs

Query patterns mined

48,210

last 30 days

Consumers tracked

312

users + service principals

PII fields tagged

PDPL / GDPR / PCI

Conversation

Ask a question about your catalog — every answer cites the tables it used.

Why usage-based ranking matters

Two tables can look identical in schema, but only one is trusted by the business.

pos_trn_dtl (retail_pos_uae)

8,412 queries / 7d · joined into 92% of Finance dashboards · owned by Retail Ops · Trusted layer.

Rank #1 — real usage confirms it's the source of truth.

pos_trn_dtl_legacy (retail_pos_uae_archive)

Schema-identical · 4 queries / 7d · last write 14 months ago · no owner.

Rank #47 — deprioritized despite matching schema. A metadata-only catalog would tie these.