"NVIDIA's Auto Ontology builds a semantic graph reasoning layer over your data."

Read that sentence and you might assume it's an open-source framework. It's not. NVIDIA Auto Ontology has no official open-source repository. It's a paid enterprise product. But an Aiven community demo reveals its actual architecture.

This article analyzes the demo source to understand Auto Ontology's real structure and compares it with conventional vector search.

Auto Ontology's actual architecture

The demo repo integrates these components:

Concern Aiven component
Chat / NL-to-SQL AI Gateway (https://ai.aiven.io/v1)
Embeddings OpenSearch ML Commons (apps/os-embed-shim)
Ontology store + vectors PostgreSQL + pgvector
Analytics source saas_analytics database (conflicting revenue definitions)

The core: a proprietary NVIDIA engine integrated with Aiven's AI Gateway, OpenSearch, and PostgreSQL. Requires NVIDIA NGC credentials and an Aiven account.

The demo scenario: "What was Q3 2025 booked revenue?"

The demo scenario clarifies what problem Auto Ontology solves:

Question: What was booked revenue in Q3 2025?
Expected answer: 717000 USD (finance invoices only)

The key: the term booked_revenue was certified in the ontology. A certified term maps to a precise SQL query:

SELECT SUM(amount) FROM finance.invoices
WHERE status = 'booked' AND quarter = 'Q3_2025'

Without the ontology, what happens?

"Revenue" means:
- Sales ARR: hundreds of thousands
- Marketing pipeline: millions
- Finance booked: 717,000

All three are "revenue." Vector search cannot distinguish them.

Vector search vs. ontology-based retrieval

Here's the structural difference in a RAG system:

Vector search (conventional)

Question: "What was last quarter's revenue?"
  → Text similarity search
  → All documents containing "revenue" get similarity scores
  → Sales, Marketing, and Finance documents all score high
  → LLM combines all results → ambiguous answer

Ontology-based retrieval (Auto Ontology)

Question: "What was last quarter's revenue?"
  → AI Gateway checks ontology-certified terms
  → "booked_revenue" is certified → routed to finance.invoices
  → "revenue" is uncertified → Sales/Marketing excluded
  → Precise SQL query → clear answer

v1 limitations: an unfinished product

The demo source reveals clear v1 limitations:

  • No embeddings endpoint: AI Gateway is chat-only. The embeddings endpoint is tracked as AIG-416 (JIRA). The demo uses an OpenSearch ML Commons shim as a workaround.
  • Rerank disabled: Disabled in v1 due to API shape mismatch. Chat works.
  • DataHub not deployed: DataHub handles estate catalog/lineage; Auto Ontology handles certified answers. They serve different purposes.
  • NGC authentication required: Docker images build from vendor/nvidia-auto-ontology; NGC app tags require authentication for public pull.

What developers should check

When evaluating Auto Ontology:

  • Not open-source: No official open-source repository exists. The Aiven demo is community-contributed.
  • Unknown pricing: Enterprise product with no publicly available pricing information.
  • Not a vector search replacement: Auto Ontology doesn't replace vector search. It supplements the area where vector search fails — "ambiguous terms."
  • Certified terms are the real value: Only certified glossary terms produce precise queries. Uncertified terms behave like regular vector search.

When vector search returns all documents containing "revenue," Auto Ontology executes a precise SQL query on finance.invoices only when "booked_revenue" is a certified term. This difference only matters when your data has clear term definitions.

The product itself is still v1 with no official open-source release. But the approach to solving "vector search limitations" is a direction worth noting for RAG system designers.