Semantic Modeling for Snowflake Cortex

Overview

Semantic Modeling bridges the gap between raw data tables and business understanding. When you connect your data sources to Precog and load them into Snowflake, Precog can automatically generate semantic models that describe your data in terms analysts and business users understand.

These semantic models power Snowflake Cortex Analyst, enabling natural language queries against your data warehouse. Instead of writing SQL, users can ask questions like "What were our top-selling products last quarter?" and get accurate answers.

Why It Matters

Data teams spend significant time translating business questions into SQL queries. Semantic models capture this translation once, making it available to everyone in the organization.

With Precog-generated semantic models:

Business users can query data using natural language through Snowflake Cortex Analyst
Analysts spend less time on ad-hoc requests and more on strategic analysis
Data engineers define meaning once rather than repeatedly explaining table structures

The result is faster insights with fewer errors and less back-and-forth between teams.

How It Works

When you set up a Precog AI Assistant, the platform analyzes the schemas across your connected sources and generates a YAML-based semantic model following Snowflake's specification. This model includes:

Logical Tables

Each source table becomes a logical table with clearly defined columns categorized as:

Dimensions — Categorical attributes for grouping and filtering (e.g., customer name, product category)
Time Dimensions — Date and timestamp fields for time-based analysis
Facts — Numeric measures that can be aggregated (e.g., order amount, quantity)

Metrics

Pre-defined calculations that combine facts with aggregation logic:

Revenue = SUM(order_amount)
Average Order Value = AVG(order_amount)
Customer Count = COUNT(DISTINCT customer_id)

Relationships

How tables connect to each other for multi-table queries:

Orders → Customers (many-to-one via customer_id)
Order Items → Products (many-to-one via product_id)

Filters

Common query constraints that business users frequently apply:

Active Customers Only
Last 12 Months
Exclude Test Transactions

The Semantic Model Format

Precog generates semantic models in YAML format, following Snowflake's Cortex Analyst specification. Here's an example structure:

name: ecommerce_semantic_model
description: Semantic model for e-commerce analytics

tables:
  - name: orders
    description: Customer orders with transaction details
    base_table:
      database: ANALYTICS
      schema: ECOMMERCE
      table: ORDERS

    dimensions:
      - name: order_status
        description: Current status of the order
        expr: ORDER_STATUS
        data_type: VARCHAR

    time_dimensions:
      - name: order_date
        description: Date when the order was placed
        expr: ORDER_DATE
        data_type: DATE

    facts:
      - name: order_amount
        description: Total amount of the order
        expr: ORDER_AMOUNT
        data_type: NUMBER

    metrics:
      - name: total_revenue
        description: Sum of all order amounts
        expr: SUM(order_amount)

    filters:
      - name: completed_orders
        description: Only completed orders
        expr: ORDER_STATUS = 'COMPLETED'

relationships:
  - name: orders_to_customers
    left_table: orders
    right_table: customers
    relationship_columns:
      - left_column: CUSTOMER_ID
        right_column: ID
    join_type: left_outer
    relationship_type: many_to_one

Working with Semantic Views in Precog

Viewing Your Semantic View

After your AI Assistant completes generation, Precog creates a Snowflake semantic view. You can check generation status from the AI Assistants page in the workspace navigation.

In Snowflake, you can find the semantic view in the schema associated with your connection.

Multi-Source Views

When your connection includes multiple data sources, Precog generates a unified semantic view that spans all of them. This enables natural language queries that join across sources — for example, correlating CRM pipeline data with accounting revenue.

Regenerating Views

Semantic views are regenerated automatically when schema changes are detected. You can also trigger regeneration by updating the use case description in your AI Assistant configuration.

Using Semantic Views with Snowflake Cortex Analyst

Precog automatically creates a Snowflake semantic view, so you can start using Cortex Analyst right away.

Users can ask questions in plain English:

"Show me monthly revenue trends for the past year"
"Which customers have the highest lifetime value?"
"Compare sales by region for Q4"

Cortex Analyst translates these questions into SQL using your semantic view's definitions.

Best Practices

Writing Effective Use Case Descriptions

The quality of your semantic model depends on clear, detailed use case descriptions. A well-written use case helps Precog generate a more accurate and useful model, which leads to better natural language query results in Cortex Analyst.

The Formula

A good use case includes four elements:

[Business Domain] + [Key Metrics/KPIs] + [Key Entities] + [Analysis Type]

When setting up your AI Assistant, take time to describe:

What business questions this data should answer
Who will use the data (analysts, executives, operations teams)
Key metrics and KPIs that matter for this data
Important relationships between entities

Examples: Poor vs. Better Use Cases

Vague descriptions produce generic models. Specific descriptions produce models tuned to your actual needs.

Accounting (Xero)

Poor	Better
"Financial reporting"	"Analyze accounts payable and receivable performance including invoice aging, payment trends, and cash flow by customer and supplier"
"Invoice analysis"	"Track invoice processing including outstanding amounts, overdue rates, and payment timing by customer and invoice type"
"Cash flow"	"Monitor cash flow including receipts, payments, and bank balance trends by account and period"

Sales/CRM (HubSpot)

Poor	Better
"Sales analysis"	"Track sales pipeline performance including deal velocity, win rates by stage, revenue by product line, and rep performance metrics"
"CRM data"	"Analyze customer engagement including contact activity, deal progression, and conversion rates by company size and industry"
"Pipeline report"	"Track deal pipeline including stage duration, drop-off rates, and forecast accuracy by owner and product"

Quick Templates

Use these templates as starting points for your use cases:

For accounting data:

Analyze [process] including [metric1], [metric2], and [metric3]
by [dimension1] and [dimension2]

Example: "Analyze accounts receivable including aging analysis, collection rates, and outstanding amounts by customer segment and invoice type"

For sales/CRM data:

Track [sales/marketing process] performance including [metric1],
[metric2], and [metric3] across [entity1] and [entity2]

Example: "Track deal pipeline performance including conversion rates, average deal size, and sales cycle length across products and territories"

Common Metrics by Domain

Include metrics relevant to your data source:

Accounting: Invoice aging, outstanding amounts, overdue rates, payment timing, days sales outstanding (DSO), cash flow, receipts, payments, spend by category, budget variance

Sales/CRM: Deal velocity, win rates, close rates, pipeline value, forecast accuracy, conversion rates by stage, average deal size, revenue by product, contact activity, engagement scores, lead source attribution

What to Avoid

Vague terms alone: "analysis" or "reporting" without specifics
Missing metrics: Not stating what you want to measure
No dimensions: Omitting how you want to slice the data
Assumed context: The system doesn't know your business jargon

Practical Insight

Semantic modeling transforms Precog from a data loading tool into an analytics enabler. By capturing business meaning at the point of data ingestion, you create a foundation for self-service analytics that scales across your organization.

Start with core business entities — customers, orders, products — and expand as users identify additional needs. Connections with multiple data sources get a unified semantic model that enables cross-source queries. The goal isn't perfect coverage on day one, but a model that grows with your analytics maturity.

Semantic View Quick Start — Quick start guide for generating your first semantic view
Semantic Layer — View your semantic model in Precog, including the ERD visualization and OSI JSON download
Knowledge Base — Curate business facts that drive semantic model generation
Semantic Modeling FAQ — Common questions about compatibility, viewing, and troubleshooting

Viewing Your Semantic Model in Precog

Once your AI Assistant generates a semantic model, you can explore it directly in Precog through the Semantic Layer page. This includes:

Summary statistics — See how many datasets, fields, relationships, and metrics your model contains
Entity relationship diagram (ERD) — An interactive visualization showing how your datasets connect to each other
Dataset details — Click into any dataset to see its fields, descriptions, and base table mapping

Precog uses the OSI (Open Semantic Interface) JSON format for semantic models. You can download the full model as a .json file from the Semantic Layer page for use with external tools, auditing, or version control.

When your Knowledge Base factors change, the Semantic Layer page alerts you so you can regenerate the model to reflect your latest business knowledge.