Holistics AI Architecture

What is Holistics AI?

Holistics AI enables end-users to get reliable analytics insights through natural language conversations—without the accuracy and reliability problems commonly faced by other AI-powered BI tools.

Foundations

Three foundational pillars enable Holistics to deliver reliable AI-assisted self-service analytics:

1. Rich Semantic Modeling Layer: Business metrics, dimensions, and relationships are defined once to provide comprehensive business context for AI.
2. Analytics Query Language (AQL): A composable, analytics-centric query language that’s built for AI reasoning. It allows AI to focus on generating high-level analytics logic instead of low-level execution details.
3. Analytics Definitions as Code: Every analytics artifact is text-based code that enables AI to easily read existing definitions and generate new ones. Comes with built-in version control and governance.

Mechanism & key differentiation

Unlike other natural-language BIs, which usually translate natural language directly to SQL, Holistics makes use of our own intermediary semantic query language called AQL (Analytical Query Language).

Based on users’ natural-language queries, the AI system first generates the AQL query. These queries are compiled into SQL before being sent off to the customer’s data warehouse.

Here are some unique attributes of AQL that are relevant for our AI-assisted self-service:

• AQL is database-agnostic, so the AI model doesn’t have to worry about the specifics of different databases (or even different database versions). Translating from AQL to SQL is handled deterministically by the AQL compiler.
• AQL comes with a library of pre-built functions that handle common analytical use cases (period comparison, percent of totals, nested aggregations, running total calculations, etc). This means the AI model doesn’t have to generate wrangled SQL logic for these use cases.
• AQL is a composable query language. That means complex operations can be broken down into smaller, modular operations and combined together (like Lego blocks).
• AQL is semantic layer-aware. This means the AI model will generate AQL that reuses existing definitions from the semantic layer, instead of coming up with its own metric formulas. Also, thanks to composability, it can derive new metrics based on existing definitions instead of starting the entire formula from scratch.

Unique benefits

Given all of the above, asking the AI’s language models to generate AQL (instead of SQL directly) gives us a few unique benefits (compared to other NLQ BI tools):

• Verifiable & readable: The generated AQL is often a lot more compact (smaller in size) and contains higher-level logic. This makes it easier for humans to comprehend and verify the AI-generated answers.
• Reliable: Because AI now operates at the high-level language, it doesn’t have to worry about lower-level details (like database-specific syntaxes, or specific SQL aerobics to achieve common analytic use cases). Hence, it should be a lot more reliable (accurate), compared to LLMs that generate SQL directly.
• Governed: Because AQL works directly with the semantic layer logic (metrics and dimensions definitions), the generated logic is ensured to use the correct definitions and goes through proper access control checking prior to serving end users.
• Capable: It can handle a lot more complex use cases thanks to composability and pre-built high-level analytical functions.