Build a semantic layer that helps agents interpret data accurately
When an agent retrieves numbers from a database but does not know what they mean, the results are unreliable. This tool creates a knowledge graph from enterprise metadata that organizes data location, meaning, and source, helping agents provide accurate answers. It validates natural language questions against business terms when converting them to SQL and displays the reasoning behind each answer.
Creating the semantic data layer
The tool reads database schemas and sample data to generate a draft ontology. This ontology includes business concepts (terms), relationships between tables, and validated SQL expressions. The generated draft undergoes a human review and certification process, where term definitions and SQL logic are finalized. The certified ontology is stored in Postgres, with embeddings stored in pgvector for vector search. This prepared layer serves as the reference standard when agents receive questions.
The agent's workflow for answering questions
The agent extracts intent and entities from the question and finds ontology candidates via vector search. It writes SQL based on the found candidates, then validates filter values and queries against the database. If validation fails, the SQL is corrected; if successful, the query is executed, and the answer is returned along with the generated SQL and processing steps. If the question asks for items not present in the data, the agent explicitly states it cannot answer and suggests alternatives. All answers include sources and validation processes, allowing results to be cross-checked with the original data.
Connecting to agents
Read-only tools are provided via an MCP server, invoked with each user's permissions. You can check question coverage with check_answerable and retrieve answers, SQL, and row data with ask_question. Using the REST API allows receiving answers via server-side event streams, authenticated by API tokens. The web UI allows managing chat, data catalog, analytics, and settings, and enables visual exploration of ontology terms and relationships. It is compatible with the Apache Ossie standard, supporting import and export of model definitions. Provides read-only tools invoked with user permissions. Receives answers via server-side event streams using API token authentication. Manages chat, data catalog, analytics, and settings, and visualizes ontology terms and relationships. Compatible with the Apache Ossie standard for importing and exporting model definitions.
Deployment and model configuration
You can run the local stack with Docker Compose or deploy to Kubernetes using Helm charts. The default configuration consists of PostgreSQL, pgvector, a FastAPI backend, and a Next.js frontend. Inference uses NVIDIA NIM endpoints, with LLM, embedding, and reranking models configured via environment variables. Each model role is defined by a set of endpoint, API key, and model name, with unset fields replaced by defaults. Database connections support ten engines, including PostgreSQL, MySQL, and Snowflake, while DuckDB and SQLite support only connection strings.
Pre-use considerations
This tool does not accept external contributions and is maintained by the NVIDIA team. It is distributed under the Apache License 2.0, with third-party components listed in separate documentation. Enabling the KumoRFM prediction feature uploads connected tables to an external service, so data review must be completed first. Ports 3001 and 3002 must be kept on a private network, and the ingest service endpoint has no self-authentication. If browser session cookies are outdated, the UI may not load correctly, requiring you to clear site data and log in again.
