Ontology & Knowledge Graph for AI Agents

An agent with search can find candidates. An agent with a knowledge graph can find the answer, and show its work. Collate exposes the graph to AI assistants through its MCP Server, so any MCP client, Claude, Cursor, VS Code, Goose, or your own agent, can traverse it directly.

Why This Changes What an Agent Can Do

Three ways an agent can look for something, in increasing order of reliability:

Approach	What it returns	Failure mode
Keyword search	Assets whose text matches	`revenue` returns fourteen tables and no basis for choosing
Semantic search	Assets whose embedding is close	The most similar table is not necessarily the governed one
Graph traversal	The exact set of nodes reachable by named edges	Returns nothing when your assumption was wrong, which is the correct behavior

The third is the only one where the agent’s answer is checkable. When an assistant says “these six dashboards depend on this column,” it can hand you the SPARQL and the triples. When it says it from memory, you cannot tell the difference between a correct answer and a fluent one. Grounding agents in a graph converts a trust problem into an auditing problem, which is a much better problem to have.

The Tools

Five MCP tools cover the graph.

Access model. All five require an administrator principal, they return 403 for ordinary users and for bot tokens. Four of them (sparql_query, entity_neighborhood, find_by_tag, shacl_validate) are not advertised at all when RDF is disabled, so a client never sees a tool it cannot call. ontology_describe is always advertised, because its default path serves the bundled ontology document from the classpath and works with RDF off.

`ontology_describe`: Start Here

Returns the OpenMetadata ontology: the full canonical document, or a SPARQL DESCRIBE for a single class or property URI.

Parameter	Notes
`resource`	Optional. Full URI of a class or property, for example `https://open-metadata.org/ontology/Column`. Omit for the whole ontology.
`format`	`turtle` (default), `jsonld`, `ntriples`, `rdfxml`.
`maxBytes`	Default and maximum 80,000; minimum 1,024.

The full ontology is roughly 65 KB, so prefer a focused DESCRIBE.

Make this the agent’s first call. An LLM asked to write SPARQL against an unfamiliar schema will invent predicate names that sound right, om:hasOwnerName, om:parentTable, and get zero rows with no error. One ontology_describe call replaces that guessing with the actual declarations. Put it in your system prompt.

`sparql_query`

Read-only SPARQL (SELECT, ASK, DESCRIBE, CONSTRUCT) against the knowledge graph.

Parameter	Notes
`query`	Required. Use `https://open-metadata.org/ontology/` as the `om:` prefix.
`format`	`SELECT`/`ASK`: `json` (default), `xml`, `csv`, `tsv`. `CONSTRUCT`/`DESCRIBE`: `turtle` (recommended), `jsonld`, `ntriples`, `rdfxml`.
`inferenceLevel`	`none` (default), `rdfs`, `owl`, `custom`.
`maxBytes`	Default and maximum 80,000; minimum 1,024.

UPDATE, INSERT, DELETE, DROP, LOAD, CLEAR, and CREATE are rejected. SERVICE clauses against external endpoints are rejected unless the target is on the federation allowlist, which is empty and disabled by default, so in a stock deployment this tool never reaches outside the local graph. Two response fields matter for agent behavior:

truncated: true with a byteCount when the body exceeds maxBytes. Narrow the query with LIMIT rather than raising maxBytes, values above the maximum are clamped, because a larger body is discarded wholesale by the response budget.
warning when the requested inferenceLevel could not be applied because the graph was too large. Results are then NOT inferred. An agent that ignores this field will confidently report an incomplete answer.

`entity_neighborhood`

The n-hop neighborhood of one entity: hasColumn, belongsToSchema, hasTag, owners, lineage, and so on. Ontology and SHACL definition graphs are excluded so schema triples do not consume result slots.

Parameter	Notes
`entityId`	Required. The entity’s `id` UUID.
`entityType`	Required. Singular, `table`, `dashboard`, `pipeline`, `glossaryTerm`, …
`depth`	1–3, default 2.
`limit`	Solution rows per traversal branch share, not a triple count. Default 200, max 2,000.

Returns the full 1..depth-hop subgraph as Turtle in triples, plus an edges array that is a flat direct (1-hop) adjacency summary of the start entity only, direction, predicate, neighbor URI, neighbor label. For second- and third-hop traversal, parse triples, not edges. Each traversal branch gets its own share of limit, so deeper hops are always represented rather than being crowded out by a high-degree start node.

`find_by_tag`

Every entity carrying a classification tag or glossary term, walking om:hasTag and om:hasGlossaryTerm.

Parameter	Notes
`tagFqn`	Required. For example `PII.Sensitive` or `BusinessTerms.Customer`.
`entityType`	Optional filter.
`limit`	Default 50, max 500.
`offset`	Default 0.

This is the “show me everything classified PII” tool, and it is a single call rather than a SPARQL round trip.

`shacl_validate`

SHACL validation of a single entity’s subgraph or the whole dataset.

Parameter	Notes
`entityId` + `entityType`	Scope to one entity.
`entityUri`	Alternative full URI.
`fullGraph`	Must be `true` to validate the entire dataset when no scope is given. Opt-in because it is expensive.
`format`	`turtle` (default) or `jsonld`.
`maxBytes`	Default and maximum 80,000; minimum 1,024.

conforms and violationCount are always returned in full even when the report body is truncated, so an agent can act on the verdict without parsing a partial report. Read-only; never blocks writes.

Complementary Tools

The graph tools sit alongside the rest of the MCP tool set. The ones that pair best:

Tool	Use it for
`semantic_search`	Meaning-based discovery when you do not know exact names. Good first step; hand results to the graph tools to confirm.
`find_context`	Semantic search over the company-knowledge layer, glossary definitions, metric definitions, Context Center notes, when no asset has been chosen yet.
`get_entity_lineage`	A compact lineage graph without writing SPARQL.
`get_entity_details`	Full entity payloads once the graph has identified them.

The pattern that works: semantic search to find candidates, graph traversal to decide between them, entity details to read the winner. Search is recall; the graph is precision.

Agent Recipes

Impact analysis before a schema change

search_metadata or get_entity_details → resolve the table and column IDs.
entity_neighborhood at depth 2 → immediate structural context.
sparql_query with prov:wasDerivedFrom+ → the full downstream set at any depth.
find_by_tag on any tier or PII tag found → who needs to be told.

Ask the agent to return the query alongside the answer. That is the artifact a reviewer checks.

Privacy review: where did this sensitive data end up?

find_by_tag with PII.Sensitive → the asserted set.
sparql_query with inferenceLevel: custom → the propagated set, if the PII-propagation rule is materialized.
Diff the two → columns that are sensitive by inference but not yet tagged. That diff is the remediation list.

Answering a business question correctly

find_context → which concept the question is actually about.
sparql_query following om:hasGlossaryTerm → the assets that realize it.
Filter by tier, owner, or contract → the governed one, not merely a matching one.

This is the flow that stops an agent from answering a revenue question off revenue_v2_final_DONOTUSE.

Catalog health audit

sparql_query → tables with no owner, no description, no glossary term.
shacl_validate with fullGraph: true → structural violations.
sparql_query → concepts with no realization (ontology debt).

Three queries, one report, no dashboard to build.

Prompting Notes

Things that measurably improve results:

Force schema discovery first. “Before writing any SPARQL, call ontology_describe for the classes involved.” Predicate hallucination is the dominant failure mode and this eliminates most of it.
Always scope the named graph. Tell the agent to wrap instance patterns in GRAPH <https://open-metadata.org/graph/knowledge> { … }. Unscoped queries return ontology triples and confuse the model about its own results.
Require the query in the answer. “Show the SPARQL you ran.” Cheap, and it makes every answer auditable.
Teach it to read truncated and warning. An agent that ignores these reports partial results as complete.
Budget with LIMIT, not maxBytes. maxBytes is clamped at 80,000; the fix for a large result is a narrower query.

Setting It Up

Enable RDF and index the graph

See RDF Knowledge Graph Indexing. Until the initial index runs, the tools return empty results.

Connect an MCP client

Follow Connect Your MCP Client. OAuth 2.0 is recommended; a Personal Access Token works where browser login is unavailable.

Authenticate as an administrator

The graph tools require an admin principal and reject bot tokens. If the tools are missing from the client’s tool list, that is the gating working, check GET /api/v1/rdf/status and your principal.

Verify

Ask the assistant to call ontology_describe with no arguments. If it returns the ontology, the wiring is correct.

Troubleshooting

The graph tools do not appear in my client

Expected when rdf.enabled is false, the four RDF-dependent tools are withheld rather than advertised and failing. Check GET /api/v1/rdf/status. ontology_describe should still be listed.

403 on every graph tool

The principal is not an administrator, or it is a bot token. Both are rejected by design.

Queries return zero rows

Usually a missing GRAPH clause or an invented predicate. Have the agent call ontology_describe and re-derive the query.

Results look incomplete

Check for truncated: true and for a warning saying inference could not be applied. Both mean the answer is partial.

Query capacity or timeout errors

The admission guard allows 8 concurrent queries globally and 2 per principal, with a 30-second timeout. Narrow the query, or materialize the traversal as an inference rule.

MCP Server

Installation, authentication, and clients.

SPARQL cookbook

Queries worth putting in an agent’s prompt.