Generates an ontology of terms, SQL attributes and relationships from database metadata, stores it in Postgres with pgvector, and uses it to answer natural-language questions with SQL and provenance. It ships a Next.js web UI, FastAPI backend and ingestion services, plus an MCP server, REST API and agent skills. Inference defaults to NVIDIA NIM endpoints; self-hosted or OpenAI-compatible endpoints can be set per model role.
Who it is for: Teams building text-to-SQL agents over enterprise databases
Strengths
Connects to ten engines including Postgres, Snowflake, Databricks, ClickHouse, DuckDB and SQLite
Each answer returns its SQL and steps; check_answerable flags questions the data cannot cover
Read-only MCP server and streaming REST API work with Cursor and Claude Desktop
Imports and exports semantic models as Apache Ossie YAML
Weaknesses
Ingestion service has no authentication; ports 3001 and 3002 must stay private
Defaults need an NVIDIA API key unless each model role points elsewhere
Project is not accepting external contributions
Quickstart needs a second repo (auto-ontology-eval) and uv to seed the sample database
What it needs
Python, NVIDIA NIM (Nemotron), OpenAI-compatible endpoints, self-hosted NIM
Needs PostgreSQL 17 with pgvector, NVIDIA NIM endpoint or OpenAI-compatible LLM, Docker Compose or Kubernetes (Helm), HashiCorp Vault (optional)
xerjElasticsearch-compatible search server that auto-indexes folders for coding agents
Live demo ↗
(opens in a new tab)19.2k stars+11,240 (+141%) in 30 daysApache-2.0