Oleksandr Kuzmenko, EditorPublished 6 min readAI-assisted · editor-reviewed
TL;DR
- Open-sourced under Apache 2.0 with native support for Snowflake, Databricks, Postgres, and DuckDB.
- Exposes relational semantic metadata to AI agents via MCP, REST API (SSE), and CLI.
- Stores credentials in HashiCorp Vault when configured, falling back to plaintext database storage if omitted.
- Integrates with NVIDIA NIM for reranking, entity extraction, and embedding generation.
Key facts
- License
- Apache 2.0
- Default Reranker
- nvidia/llama-nemotron-rerank-vl-1b-v2
- Supported Databases
- Postgres, Databricks, Snowflake, DuckDB
- Protocol Interfaces
- MCP, CLI, SSE REST API
Architecture and Supported Databases
Auto-Ontology operates as an intermediate grounding layer between raw tabular data and language models. The open-source project supports four primary database engines: Databricks, PostgreSQL, Snowflake, and DuckDB. Rather than relying on standard text-to-SQL prompt injection, it ingests enterprise metadata, generates an entity-relationship knowledge graph, and tracks lineage.
Connection credentials can be stored securely through native HashiCorp Vault integration (configured via VAULT_ADDR, VAULT_NAMESPACE, VAULT_ROLE_ID, and VAULT_SECRET_ID). If Vault configuration is omitted or incomplete, credentials added through the UI are stored unencrypted in Postgres catalog rows. User authentication supports Single Sign-On (SSO) and RBAC powered by Better Auth, with scoped API tokens that carry identical execution permissions to the issuing user identity.
Model Integrations and Endpoints
The stack utilizes NVIDIA NIM inference microservices for routing and embedding generation. By default, it configures nvidia/llama-nemotron-rerank-vl-1b-v2 for context reranking alongside dedicated text-embedding and non-reasoning extraction models via environment variable triplets (DEFAULT_MODELS_ENDPOINT, DEFAULT_MODELS_MODEL, DEFAULT_MODELS_API_KEY). Public client calls are exposed through Server-Sent Events (SSE) at POST /api/chat/completions.
Deployment via Docker
Developers can deploy the stack locally with Docker Compose v2. It provisions the Next.js UI on port 3000, the Python API gateway on port 3001, an ingestion worker, and an isolated PostgreSQL instance for semantic metadata.
Try it in 2 minutes
git clone <repo-url> auto_ontology && cd auto_ontology
# Fill in .env with AUTH_SECRET, APP_URL, and DEFAULT_MODELS_API_KEY
docker compose up -dWhen to use
- When grounding LLM agents on relational systems such as Snowflake, Databricks, or Postgres.
- When serving enterprise metadata to external agents via Model Context Protocol or REST APIs.
When NOT to use
- When running strictly unstructured document retrieval without tabular or relational databases.
- When deploying in environments where HashiCorp Vault is required but cannot be configured, as credentials will sit in plaintext.
What to do today
- Audit existing internal database connectors to configure HashiCorp Vault variables before deploying Auto-Ontology to production.
- Test Model Context Protocol endpoints to connect local agents with curated enterprise schemas.
Step names are taken from What to do today.
Step names are taken from What to do today. Audit existing internal database connectors to configure HashiCorp Vault variables before deploying Auto-Ontology to production.. Test Model Context Protocol endpoints to connect local agents with curated enterprise schemas.
- Audit existing internal database connectors to configure HashiCorp Vault variables before deploying Auto-Ontology to production.
- Test Model Context Protocol endpoints to connect local agents with curated enterprise schemas.