Skip to content
← Services·Implementation · Databricks Genie
02 · AI GENIE

Ask your warehouse in plain English.

A defensible Databricks Genie deployment — every answer comes from a certified gold layer, ships with the SQL it ran, respects Unity Catalog row + column masks, and is gated by an eval suite before any change reaches users.

01 / 06Why Genie

Most NL-to-SQL tools are demo-good, not deploy-good.

Generic NL-to-SQL chatbots are easy to demo and hard to defend. They guess schemas, ignore business definitions, hide the SQL they ran, and answer the same question two different ways on consecutive Mondays.

The defensible version is Genie pointed at a certified gold layer, with a metrics + glossary semantic surface, full SQL transparency, per-user identity, and an eval suite that gates every change. The work isn’t in the chatbot — it’s in the data layer underneath. That’s what we build.

02 / 06Architecture

The shape of a trustworthy Genie.

Questions come in from web, Slack, Teams, or API. Genie plans against the certified gold layer + glossary, generates editable SQL, runs it on serverless DBSQL, and returns a cited answer. Unity Catalog governs the lot — identity, masks, lineage, audit, and per-query cost.

Reference · Databricks Genie · NL → SQL
User surfacesWeb appEMBEDDED CHATSlack · TeamsCONVERSATIONALREST APISERVER-TO-SERVERDatabricks workspaceNATIVE GENIE SPACEGenie spaceGenieNL → SQL · AGENTQ · NATURAL LANGUAGE"Top claims by region last quarter"Plan · table choice · joinsSQL · transparent · editableEvals · accuracy gatesFollow-ups · session memoryCertified dataGold tablesCERTIFIED · UCDOMAIN-OWNEDMetrics layerDBT · SEMANTICCANONICAL DEFSGlossary · synonymsPROMPT-INJECTEDComputeDBSQL warehouseSERVERLESS · ELASTICPHOTON · CACHEDMaterialized viewsINCREMENTAL · FASTResult cacheSUB-SECOND REPLAYAnswerChart · table · KPIRICH RESPONSECited SQLSHOW · COPY · EDITFollow-up suggestionsDRILL-DOWN · COMPAREUnity Catalog · Governance & TrustRow · column · view masksLineage · query → tablesAudit · every QEntra ID · per-user identityCost attribution · per QEval suite · accuracy gatesFlowUSERGENIECERTIFIEDCOMPUTEANSWERQuestion in plain English · Genie plans against certified gold tables · SQL on serverless DBSQL · cited answer with chart + drill-down
User surfaces Genie agent Certified data DBSQL compute Answer surface Unity Catalog
03 / 06What we build

Six capabilities that make it defensible.

01

Certified gold tables

Genie answers only from the data you’ve certified. We build the gold layer first so analysts and business users get answers from one canonical source — not whichever table is closest.

02

Metrics + glossary

A semantic layer with canonical metric definitions, synonyms, and business glossary prompts. "Active claim", "FY", "open period" mean the same thing every time.

03

NL → SQL with citations

Every answer shows the SQL behind it, the tables it touched, and the rows it returned. Editable, copyable, defensible — never a black box.

04

Multi-surface delivery

Native Databricks Genie space, Slack and Teams integration, embeddable web widget, and a server-to-server REST API for tooling.

05

Unity Catalog governance

Per-user identity via Entra ID. Row and column masks apply. Every question and answer is audited. Cost attribution per query.

06

Eval suite

A pinned set of questions with known-good answers, run on every change. Accuracy gates before promoting any model or schema change to production.

04 / 06Who it's for

Built for you if…

  • 01

    You have a governed lakehouse already (Databricks, Fabric, Snowflake) and want the answers in plain English without the gold layer turning into spreadsheet exports.

  • 02

    Your analysts spend more time answering 'can you pull X for me?' questions than building real models — and you'd like to give that work back.

  • 03

    You need every answer to be defensible: who asked, what SQL ran, which rows, when. Not a black-box chat with the warehouse.

  • 04

    You're cautious about NL-to-SQL accuracy and want a real eval framework that gates changes, not vibes.

  • 05

    Australian-hosted, identity-aware, with usage and cost visible per user and per query.

05 / 06How we deliver

Data first, chatbot second.

We do the unglamorous work first — certify the gold layer, define the metrics, write the glossary — then stand up Genie on top.

  1. — 01

    Certify the gold layer

    We start with the data, not the interface. The first 2–3 cycles harden a small certified gold layer with metrics + glossary so Genie has somewhere honest to plan against.

  2. — 02

    Stand up the Genie space

    Genie space scoped to those certified tables, glossary loaded, eval suite seeded with the questions that actually matter to your business.

  3. — 03

    Roll out + tune

    Beta inside a small team, tune prompts and synonyms with real questions, then expand to Slack / Teams / embedded surfaces. Eval suite gates every change after launch.

06 / 06FAQ

Common questions.

Start a conversation

Conversational data, done defensibly.

Send a brief and we’ll reply within one business day with a scoped response — and only take it on if we can ship something you’ll trust.