Natural Language Analytics with Snowflake Cortex for Self-Service BI

gitlab · Tech · 2026 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Self-service business intelligence

GitLab needed non-technical users to query structured data in natural language without increasing the analytics backlog or bypassing enterprise governance.

Constraints:

The system needed to support simple and complex business queries.
Early proof-of-concept accuracy was only 60%.
Data access had to respect role-based controls.
Business users needed faster time-to-insight than analyst queues could provide.

Architecture

GitLab used Snowflake Cortex with semantic models, prompt engineering, verified query feedback loops, and Snowflake role-based access controls to create governed conversational analytics.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  business-user-interface["Business User Interface"]
  snowflake-cortex["Snowflake Cortex"]
  semantic-models["Semantic Models"]
  verified-query-feedback-loops["Verified Query Feedback Loops"]
  snowflake-governance["Snowflake Governance"]
  business-user-interface --> snowflake-cortex
  snowflake-cortex --> semantic-models
  snowflake-cortex --> snowflake-governance
  semantic-models --> snowflake-cortex
  verified-query-feedback-loops --> snowflake-cortex
  snowflake-governance --> business-user-interface

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Business User Interface`	Accepts natural-language analytics questions.	`Snowflake Cortex`
`Snowflake Cortex`	Translates natural-language questions into governed analytics queries.	`Semantic Models`, `Snowflake Governance`
`Semantic Models`	Describe business data so the model can map user questions to structured queries.	`Snowflake Cortex`
`Verified Query Feedback Loops`	Improve accuracy using reviewed successful queries.	`Snowflake Cortex`
`Snowflake Governance`	Applies role-based access controls to analytics access.	`Business User Interface`

Data flow

A user asks a natural-language BI question, Cortex interprets it through semantic models and prompt context, verified-query feedback improves future behavior, and Snowflake governance enforces role-based access to results.

“utilizing semantic models, prompt engineering, verified query feedback loops, and role-based access controls.”summary, stated

“democratized data access while maintaining enterprise-grade security through Snowflake's native governance features.”summary, stated

Why this architecture

Keeping the assistant inside Snowflake let GitLab improve self-service accuracy while inheriting native governance and access controls.

“enterprise-grade security through Snowflake's native governance features.”summary, stated

“semantic models, prompt engineering, verified query feedback loops”summary, stated

Models used: Snowflake Cortex

Scale

Accuracy improved from 60% in the proof of concept to 85-95% for simple production queries.
“evolved from a basic proof-of-concept with 60% accuracy to a production system achieving 85-95% accuracy for simple queries”summary, stated
Complex queries reached 75% accuracy.
“75% for complex queries”summary, stated
Some teams reduced analytics requests by about 50%.
“reduced analytics requests by approximately 50% for some teams”summary, stated
Time-to-insight dropped from weeks to seconds.
“decreased time-to-insight from weeks to seconds”summary, stated

Alternatives considered (roads not taken)

Traditional analyst-mediated BI

Rejected because: It created backlog and delayed business answers.

“eliminating the traditional dependency on data analysts and reducing analytics backlog.”summary, stated

Ungoverned natural-language data access

Rejected because: Enterprise security required native role-based access controls.

“role-based access controls”summary, stated

Patterns

`query-rewriting` (query-generation)

Natural-language BI questions are rewritten into structured Snowflake analytics queries.

“enable non-technical business users to query structured data using natural language”summary, stated

`eval-driven-development` (quality)

Verified-query feedback loops improved production query accuracy after a weaker proof of concept.

“verified query feedback loops”summary, stated

`safety-policy-enforcement` (governance)

Snowflake role-based access controls enforce enterprise data governance.

“maintaining enterprise-grade security through Snowflake's native governance features.”summary, stated

Tradeoffs

Use Snowflake-native conversational analytics

Chose Snowflake Cortex and governance over A separate external BI chatbot.

Rationale: Keeping the system in Snowflake preserved governance while enabling natural-language access.

“using Snowflake Cortex to enable non-technical business users to query structured data using natural language”summary, stated

Use feedback loops

Chose Verified query feedback over Static prompt-only behavior.

Rationale: Production accuracy needed to improve from the POC baseline.

“verified query feedback loops”summary, stated

Gotchas (what bit them)

Proof-of-concept accuracy is not enough for self-service BI.

Root cause: Early natural-language analytics produced only 60% accuracy.

Resolution: Add semantic models, prompt engineering, and verified query feedback loops.

“evolved from a basic proof-of-concept with 60% accuracy to a production system achieving 85-95% accuracy for simple queries”summary, stated

Democratized analytics can become a governance risk.

Root cause: Non-technical users directly query structured enterprise data.

Resolution: Use Snowflake native governance and role-based access controls.

“enterprise-grade security through Snowflake's native governance features.”summary, stated

Evaluation

GitLab measured accuracy by query complexity, request reduction, and time-to-insight.

Metrics: Simple-query accuracy · Complex-query accuracy · Analytics request reduction · Time-to-insight

“85-95% accuracy for simple queries and 75% for complex queries”summary, stated

Outcomes

Simple-query accuracy reached 85-95%.
Complex-query accuracy reached 75%.
Some teams reduced analytics requests by about 50%.
Time-to-insight dropped from weeks to seconds.

The case reports team-level request reductions and production accuracy bands rather than total user volume.

“The implementation reduced analytics requests by approximately 50% for some teams, decreased time-to-insight from weeks to seconds”summary, stated