Skip to main content

​
Ibex Agent Engine

The ibex-agent-engine is the operational brain powering the AI functionality across the data platform. Built in Node.js, it executes multi-step reasoning pipelines that combine Large Language Model actions with physical SQL execution.

​
Architectural Role

This service runs strictly inside the internal Docker ajna network on port 3000. It is not exposed to the internet via Traefik. It acts exclusively as a secure worker invoked by the ibex-ai-service.

​
High-Level Working

Each agent running on the engine is defined as an execution pipeline composed of discrete tool nodes. A standard workflow to answer a user’s analytical question looks like this:
  1. Information Retrieval (tool-kb-1): The agent engine searches the RAG knowledge base for context regarding table schemas and business documentation.
  2. SQL Generation & Execution (tool-exec-1): The engine calls out to OpenRouter (https://openrouter.ai/api/v1) using OPENROUTER_API_KEY to generate the SQL query. It then routes this physical SQL to the internal ibex-analytics-service and awaits the results.
  3. Natural Language Generation (tool-gen-1): The engine makes a final LLM call to format the raw database records into a conversational response.
The pre-defined standard agent, demo_data_analyst, is hosted here, bridging the gap between natural language requests from the BI UI and DuckDB federated query results.