Skip to main content
Version: PromptQL

Architecture

Introduction​

Below, you'll find a functional overview of PromptQL's architecture and the key components that power it. At the heart of PromptQL is the semantic metadata layer — the intelligent foundation that makes natural language data queries possible.

High-level overview​

How it works​

Request initiation​

When you ask PromptQL a question, the Playground Server receives your natural language query and combines it with your semantic metadata — the intelligent layer that describes your connected data sources, schemas, relationships, business logic, and access controls.

This metadata layer is generated automatically when you connect a data source and evolves as you and your team use PromptQL.

Query plan generation​

The Playground Server sends both your question and the semantic metadata to your chosen LLM. This semantic context is what transforms a general-purpose LLM into a data expert for your specific organization. The metadata tells the LLM exactly what data exists, how it's structured, what different fields mean in your business context, and how tables relate to each other.

With this rich context, the LLM generates a precise query plan that maps your natural language request to the right data operations.

Query execution​

The Playground Server passes the query plan to PromptQL's Python runtime, which executes it through the distributed query engine. The engine uses data connectors to fetch data from your data sources, respecting all access controls and mappings defined in your metadata.

Response delivery​

The retrieved data flows back through the same path: from data sources through connectors, engine, and runtime, where it's packaged and returned to you via the Playground Server.

This architecture ensures that your LLM has complete context about your data landscape while maintaining security and performance through the distributed execution layer.

Next steps​

With a broad perspective of the different services powering a PromptQL application, we recommend learning more about the semantic metadata layer next.