> ## Documentation Index
> Fetch the complete documentation index at: https://docs.open-metadata.org/llms.txt
> Use this file to discover all available pages before exploring further.

# Enabling Ontology & Knowledge Graph | OpenMetadata

> What works out of the box, what needs a triple store, and the full configuration surface for RDF, inference, SHACL, federation, and the Ontology Studio AI assistant.

# Enabling Ontology & Knowledge Graph

Ontology and the knowledge graph are two different things with two different requirements. Knowing which is which saves a lot of unnecessary infrastructure.

## What Works with No Extra Infrastructure

The **ontology** is stored relationally. Everything below works on a stock OpenMetadata deployment:

* Ontology Studio **View** mode — graph, tree, filters, health, inspector
* Ontology Studio **Edit** mode — concepts, typed relations, attributes, axioms, mappings, patterns, subsets, merges, change sets, impact previews
* **Relationship type** management
* **Ontology Library** pack installation
* Ontology **import and export** (OWL, Turtle, N-Triples, JSON-LD, SKOS, CSV)
* **Glossary-scoped SPARQL** (`POST /api/v1/glossaries/{id}/sparql`)
* The `ontology_describe` MCP tool

## What Needs the RDF Triple Store

* Ontology Studio **Query** mode and the standalone **SPARQL Playground**
* The **3D Knowledge Graph** tab on tables
* Catalog-wide SPARQL, graph exploration, and graph export
* **Inference** and **SHACL validation** over the catalog
* **Graph insights** — importance, centrality, communities, paths, recommendations, tag analytics
* The `sparql_query`, `entity_neighborhood`, `find_by_tag`, and `shacl_validate` **MCP tools**
* **Linked Open Data** IRI dereferencing

## What Needs an LLM Provider

* Ontology Studio **AI** mode — relationship, mapping, SPARQL, and domain-draft proposals

Requires `RDF_ASK_COLLATE_ENABLED=true` **and** a configured LLM provider.

## Turning on RDF

<Warning>
  The RDF knowledge graph is **Beta** and **disabled by default**. Turning it on for a large existing catalog adds meaningful load while the initial index runs. Test in a non-production environment first and schedule the initial index for a low-traffic window.
</Warning>

Full deployment instructions — Docker, Kubernetes, Helm, sizing, and troubleshooting — are in [Enable RDF (Knowledge Graph)](/v2.1.x-SNAPSHOT/deployment/rdf-knowledge-graph). The short version:

<Steps>
  <Step title="Run Apache Jena Fuseki">
    Reachable from the OpenMetadata server. The repository ships one at `docker/rdf-store`.
  </Step>

  <Step title="Set the environment variables on both containers">
    The **migrations container** loads the ontology and shapes graphs and registers `RdfIndexApp`; the **server container** does everything else. If only one has RDF configured, installation fails with `RdfRepository not initialized`.

    ```bash theme={null}
    RDF_ENABLED=true
    RDF_STORAGE_TYPE=FUSEKI
    RDF_ENDPOINT=http://fuseki:3030/openmetadata
    RDF_DATASET=openmetadata
    RDF_REMOTE_USERNAME=admin
    RDF_REMOTE_PASSWORD=<password>
    RDF_BASE_URI=https://open-metadata.org/
    ```
  </Step>

  <Step title="Run the initial index">
    **Settings → Applications → RDF Knowledge Graph Indexing → Run Now**, or:

    ```bash theme={null}
    curl -X POST "$OM_HOST/api/v1/apps/trigger/RdfIndexApp" -H "Authorization: Bearer $TOKEN"
    ```
  </Step>

  <Step title="Verify">
    ```bash theme={null}
    curl "$OM_HOST/api/v1/rdf/status" -H "Authorization: Bearer $TOKEN"
    ```

    Expect `enabled: true` and `projectionState: READY`. `REBUILDING` means the index is still running; `DEGRADED` means something needs attention.
  </Step>
</Steps>

### Trying It Locally

The repository ships a startup script that brings up the whole stack with RDF and Fuseki enabled:

```bash theme={null}
./docker/run_local_docker_rdf.sh -m ui -d mysql -f true
```

It sets sensible development defaults, including materialized inference and `CUSTOM` as the default inference level.

## Configuration Reference

All settings live under `rdf:` in `openmetadata.yaml` and are overridable by environment variable.

### Core

| Variable                                      | Default                              | Purpose                                                                                           |
| --------------------------------------------- | ------------------------------------ | ------------------------------------------------------------------------------------------------- |
| `RDF_ENABLED`                                 | `false`                              | Master switch. Every RDF code path is a no-op until true.                                         |
| `RDF_STORAGE_TYPE`                            | `FUSEKI`                             | Backend. `QLEVER` is defined in the schema but **not implemented** — setting it fails at startup. |
| `RDF_ENDPOINT`                                | `http://localhost:3030/openmetadata` | Full URL ending in the dataset name. `RDF_REMOTE_ENDPOINT` is a deprecated fallback.              |
| `RDF_DATASET`                                 | `openmetadata`                       | Dataset name. Must match the path segment in `RDF_ENDPOINT`.                                      |
| `RDF_REMOTE_USERNAME` / `RDF_REMOTE_PASSWORD` | `admin` / `admin`                    | Basic-auth credentials.                                                                           |
| `RDF_BASE_URI`                                | `https://open-metadata.org/`         | Base URI for minted resource IRIs.                                                                |

### Connection and Write Tuning

| Variable                                  | Default | Purpose                                                |
| ----------------------------------------- | ------- | ------------------------------------------------------ |
| `RDF_CONNECT_TIMEOUT_MS`                  | `2000`  | Connection timeout.                                    |
| `RDF_REQUEST_TIMEOUT_MS`                  | `60000` | Per-request timeout.                                   |
| `RDF_WRITE_MAX_RETRIES`                   | `2`     | Retries for idempotent writes after the first attempt. |
| `RDF_WRITE_RETRY_INITIAL_BACKOFF_MS`      | `250`   | Initial retry backoff.                                 |
| `RDF_WRITE_RETRY_MAX_BACKOFF_MS`          | `2000`  | Maximum retry backoff.                                 |
| `RDF_BULK_ENTITY_BATCH_SIZE`              | `100`   | Entity models per bulk write.                          |
| `RDF_BULK_RELATIONSHIP_SOURCE_BATCH_SIZE` | `100`   | Relationship sources per bulk write.                   |
| `RDF_BULK_LINEAGE_EDGE_BATCH_SIZE`        | `50`    | Lineage edges per SPARQL update.                       |

### Inference

| Variable                              | Default  | Purpose                                                                                                                |
| ------------------------------------- | -------- | ---------------------------------------------------------------------------------------------------------------------- |
| `RDF_INFERENCE_ENABLED`               | `false`  | Enable inference on SPARQL queries.                                                                                    |
| `RDF_DEFAULT_INFERENCE_LEVEL`         | `NONE`   | `NONE`, `RDFS`, `OWL_LITE`, `OWL_DL`, or `CUSTOM`.                                                                     |
| `RDF_MATERIALIZED_INFERENCE_ENABLED`  | `false`  | Use durable per-rule inferred named graphs instead of in-memory Jena models. **Recommended past a small catalog.**     |
| `RDF_MAX_IN_MEMORY_INFERENCE_TRIPLES` | `100000` | Above this store size, inference requests fall back to direct execution — with a `warning`, and **without** inference. |
| `RDF_CACHE_INFERRED_TRIPLES`          | `false`  | Cache bounded in-memory inference models for 60 seconds.                                                               |

### Validation and Profile

| Variable                    | Default  | Purpose                                                            |
| --------------------------- | -------- | ------------------------------------------------------------------ |
| `RDF_SHACL_VALIDATION_MODE` | `REPORT` | `OFF`, `REPORT`, or `ENFORCE_IMPORTS`. Never blocks entity writes. |
| `RDF_STRICT_OWL_PROFILE`    | `true`   | Reject authored axioms outside the supported OWL 2 DL profile.     |

### Linked Data, AI, Federation

| Variable                           | Default | Purpose                                                                                |
| ---------------------------------- | ------- | -------------------------------------------------------------------------------------- |
| `RDF_DEREFERENCEABLE_IRIS`         | `false` | Authenticated content-negotiated redirects for minted IRIs (`/api/v1/lod/entity/...`). |
| `RDF_ASK_COLLATE_ENABLED`          | `false` | Expose the Ontology Studio AI mode. Manual authoring is unaffected.                    |
| `RDF_FEDERATION_ENABLED`           | `false` | Master switch for `SERVICE` clauses to external SPARQL endpoints.                      |
| `RDF_FEDERATION_ALLOWED_ENDPOINTS` | `[]`    | Allowlisted external endpoint URIs. Compared **verbatim** — trailing slashes matter.   |

<Warning>
  **Federation is a data-exfiltration surface.** A `SERVICE` clause sends parts of your query — and potentially your data — to a third-party endpoint. It is disabled by default, and the allowlist is empty. Enable it only for endpoints you control or explicitly trust.
</Warning>

### Indexing Application

`RdfIndexApp` configuration (**Settings → Applications**):

| Setting                               | Default              | Purpose                                                        |
| ------------------------------------- | -------------------- | -------------------------------------------------------------- |
| `entities`                            | `[]` (all supported) | Restrict the reindex to specific entity types.                 |
| `recreateIndex`                       | `true`               | Clear the store before indexing.                               |
| `batchSize`                           | `100`                | Entities per batch.                                            |
| `producerThreads` / `consumerThreads` | `2` / `3`            | Non-distributed reindex parallelism.                           |
| `queueSize`                           | `5000`               | Internal queue size for non-distributed reindexing.            |
| `useDistributedIndexing`              | `true`               | Partition-coordinated indexing across servers.                 |
| `partitionSize`                       | `10000`              | Entities per partition. Smaller values create more partitions. |

Only one reindex job can be active per cluster at a time.

## Verifying the Whole Thing

```bash theme={null}
# 1. RDF is on and the projection is ready
curl "$OM_HOST/api/v1/rdf/status" -H "Authorization: Bearer $TOKEN"

# 2. The graph has data
curl -X POST "$OM_HOST/api/v1/rdf/sparql" \
  -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
  -d '{"query":"SELECT (COUNT(*) AS ?n) WHERE { GRAPH <https://open-metadata.org/graph/knowledge> { ?s ?p ?o } }"}'

# 3. Inference rules are registered
curl "$OM_HOST/api/v1/rdf/rules" -H "Authorization: Bearer $TOKEN"

# 4. Shapes are loaded and the graph conforms
curl "$OM_HOST/api/v1/rdf/validate" -H "Authorization: Bearer $TOKEN" -D - -o /dev/null | grep OM-SHACL-Conforms
```

In the UI: **Govern → Ontology Studio** should show a **Query** tab, and any table should show a **Knowledge Graph** tab.

## Next

<CardGroup cols={2}>
  <Card title="Deployment guide" href="/v2.1.x-SNAPSHOT/deployment/rdf-knowledge-graph">
    Docker, Kubernetes, Helm, sizing, and troubleshooting.
  </Card>

  <Card title="Knowledge Graph" href="/v2.1.x-SNAPSHOT/how-to-guides/ontology/knowledge-graph">
    What you get once it is on.
  </Card>
</CardGroup>