Fabriciq Ontology Consumption CLI

Update Check — ONCE PER SESSION (mandatory) The first time this skill is used in a session, run the check-updates skill before proceeding.

• GitHub Copilot CLI / VS Code: invoke the check-updates skill (e.g., /fabric-skills:check-updates).
• Claude Code / Cowork / Cursor / Windsurf / Codex: read the local package.json version, then compare against remote via git fetch origin main --quiet && git show origin/main:package.json (or the GitHub API). If remote is newer, show the changelog and update instructions.
• Skip if the check was already performed earlier in this session.

CRITICAL NOTES

1. Ontology is preview. The item type value is Ontology. Wire format and limitations may change; validate against the current docs before production use.
2. This skill is read-only. It never calls createItem or updateDefinition. For schema changes, delegate to fabriciq-ontology-authoring-cli.
3. This skill does not query source data directly. It enumerates ontology grounding context, then delegates the actual data read to the per-datasource consumption skill that matches the binding source kind (see Query Routing).
4. Projections (a semantic query layer over ontology entities) are not yet GA. Until they ship, all data queries run against the source table (LakehouseTable or KustoTable) using the columns declared in the binding's propertyBindings[].
5. To find the workspace details (including its ID) from workspace name: list all workspaces and use JMESPath filtering.
6. To find the ontology item ID from workspace ID and item name: list all items of type Ontology in that workspace and use JMESPath filtering.

fabriciq-ontology-consumption-cli — Fabric Ontology Consumption via CLI

Table of Contents

Ontology Consumption Concepts

A Fabric Ontology item carries its schema as a tree of JSON files inside the item definition (same shape as authoring). Get Item Definition returns the parts as base64-encoded payloads; the consumption flow is always: fetch → decode → parse → ground → delegate.

Grounding context = the flattened, agent-ready projection of that tree: a JSON summary an LLM can read to decide which entity type to query, which column holds the key, which table to hit, and which consumption skill to invoke. Full shape + extraction recipe: grounding-extraction.md.

Property valueType allowed values (exact): String, Boolean, DateTime, Object, BigInt, Double. Integers are BigInt (not Int64); GUIDs are modelled as String. See ONTOLOGY-AUTHORING-CORE.md § EntityTypeProperty for the full source-column → valueType mapping.

Tool Stack

Ontology consumption uses the same Fabric control-plane tool stack as every other CLI skill — see COMMON-CLI.md § Tool Selection Rationale for the canonical list (install commands, prerequisite checks, base64 helpers, JSON tooling) and COMMON-CLI.md § Authentication Recipes for az login + token acquisition.

Per-datasource reads are delegated — you do not need the Kusto, Spark, or SQL CLI tools installed to use this skill for enumeration. They are only needed if you also invoke the downstream consumption skill in the same session.

Connection

Ontology consumption targets the Fabric control plane. You need the workspace ID and the ontology item ID; everything else (entity types, bindings, source tables, cluster URIs) is recovered by decoding the definition.

• Sign in + acquire the Fabric control-plane token → COMMON-CLI.md § Authentication Recipes (always --resource https://api.fabric.microsoft.com).
• Resolve workspace ID by displayName and the ontology item ID via GET /v1/workspaces/{WS_ID}/items?type=Ontology filtered by displayName → COMMON-CLI.md § Finding Workspaces and Items in Fabric (covers pagination + JMESPath filtering).
• Generic az rest invocation template → COMMON-CLI.md § Fabric Control-Plane API via az rest.

Fetching the definition (Ontology-preview LRO gotcha)

Get Item Definition is Long-Running-Operation-capable. Depending on tenant/SKU, the POST may return the definition envelope inline (200 OK) or return 202 Accepted with an x-ms-operation-id header; for the 202 case, poll https://api.fabric.microsoft.com/v1/operations/{operationId} until Succeeded, then GET …/operations/{operationId}/result to receive the parts[] array. The generic LRO recipe (capture x-ms-operation-id, poll, fetch result) is in COMMON-CLI.md § Long-Running Operations (LRO) Pattern.

Ontology-preview gotcha — prefer polling the operations/{id} endpoint over the Location header. The public Fabric LRO contract supports either, but on this Ontology LRO the Location header has been observed redirecting to an *.analysis.windows.net host; polling it with a Fabric-audience token is flaky (intermittent 401/403). Poll https://api.fabric.microsoft.com/v1/operations/{operationId} on the Fabric host instead. If you must follow Location, use the audience required by that URL. If a poll ever returns a non-2xx, read the operation .error and stop — never blind-retry the POST.

The full fetch-and-decode flow for this skill (LRO capture + part decode + tree reconstruction, Bash + PowerShell + Python helpers) lives in grounding-extraction.md § Fetch and Decode. The sibling fabriciq-ontology-authoring-cli documents the same redirect-host workaround in its LRO Header Capture section — keep the two in sync if you change one.

Source-data connections (delegated)

Running the data query itself (KQL / Spark SQL / T-SQL) uses the connection patterns owned by the sibling consumption skills:

• Eventhouse (KustoTable bindings) → EVENTHOUSE-CONSUMPTION-CORE.md § Connection Fundamentals + eventhouse-consumption-cli
• Lakehouse (LakehouseTable bindings) → sqldw-consumption-cli (default, SQL analytics endpoint) or spark-consumption-cli (only when user explicitly wants PySpark / DataFrame work)
• Warehouse bindings → ❓ not documented in the current shared ontology schema (LakehouseTable and KustoTable only); if encountered, surface to user rather than silently assuming sqldw-consumption-cli

The clusterUri, databaseName, workspaceId, and itemId your delegate needs are all already inside the decoded binding payload — do not rediscover them via the item APIs.

Consumption Scope

Until projections ship, all data reads are source queries — they run against the physical LakehouseTable or KustoTable using the columns listed in propertyBindings[]. Semantic features (inferred joins, derived measures) that require projections are not available from this skill; flag them to the user and proceed with source-level filtering.

Grounding Context

Deep recipe + full JSON shape: grounding-extraction.md. Quick index:

Grounding JSON contract — the authoritative shape lives in grounding-extraction.md § Grounding Summary Schema. Routing decisions read these fields per binding:

• source.kind (LakehouseTable | KustoTable) — picks the delegate family.
• source.workspaceId + source.itemId — read from the binding, not the ontology (cross-workspace bindings are legal).
• source.sourceSchema + source.sourceTableName — Lakehouse-only has schema.
• source.clusterUri + source.databaseName — Kusto-only.
• dataBindingType (NonTimeSeries | TimeSeries) + timestampColumnName — required for TS routing.
• propertyBindings[].sourceColumnName — the ontology-property → physical-column remap; applied before any query is composed.
• relationshipTypes[].contextualizations[].sourceKeyRefBindings[] / targetKeyRefBindings[] — arrays; composite keys are legal. Join on all entries.

Hand a trimmed subset of this JSON (not the raw base64 definition.parts[]) to downstream skills.

Query Routing

Deep recipe + per-skill invocation templates: routing.md. Quick decision table:

Invocation contract when handing off — this skill resolves the source metadata and composes the query in the target dialect, then hands both to the delegate. The delegate (sibling skill) owns the actual connection + execution.

• Eventhouse delegate → resolved connection (clusterUri, databaseName) + composed KQL text targeting sourceTableName with timestampColumnName filter + key predicate.
• SQL endpoint delegate (default for Lakehouse) → resolved connection (workspaceId, itemId, sourceSchema) + composed T-SQL text.
• Spark delegate (alternate for Lakehouse) → same resolved connection + composed Spark SQL text — use Spark-native time functions (current_timestamp() - INTERVAL 1 HOUR); do not emit T-SQL DATEADD / SYSUTCDATETIME to Spark.
• Always translate ontology property names → propertyBindings[].sourceColumnName inside the composed query text. The delegate sees only physical columns and will not resolve ontology identifiers.
• Always pass composite sourceKeyRefBindings[] / targetKeyRefBindings[] — join on every element, not just the first.

Because projections are not yet GA (see Critical Note #4), always translate the ontology property names the user mentions back into physical source column names (via propertyBindings[]) before composing the delegate query. Do not pass ontology property names to the sibling skill — it will not resolve them.

Must / Prefer / Avoid / Troubleshooting

Must

• Require explicit ontology context before routing here — the prompt must mention an "ontology" (or reference an ontology item by ID/name). Generic "Fabric IQ" or report/dataset prompts without ontology context are not ontology tasks; defer those to the matching data skill (e.g. powerbi-consumption-cli for Power BI reports). This keeps the shared "Fabric IQ" brand from over-triggering this skill.
• Clarify before routing ambiguous prompts — if the user asks "show me aircraft readings" and multiple entity types bind to aircraft-like tables, ask which entity type / binding to use. Silent guessing produces wrong data.
• Resolve WS_ID and ONT_ID before fetching the definition — hardcoded GUIDs are a top failure mode.
• Follow the LRO pattern on getDefinition — a 202 with x-ms-operation-id is normal; do not treat it as success. Poll operations/{operationId} until Succeeded, then GET …/operations/{operationId}/result. Prefer operations/{operationId} over raw Location polling (analysis.windows.net redirect can be flaky with a Fabric-audience token); if you follow Location, use the audience required by that URL. On a Failed/non-2xx poll, read .error and stop — never blind-retry.
• Decode every relevant part before answering — never respond from a cached partial view of the ontology. The caller may have added / altered entity types since you last read.
• Translate ontology property names → source column names via propertyBindings[] before generating any KQL / Spark SQL / T-SQL. The sibling consumption skills see only physical columns.
• Respect the binding type — TimeSeries requires a time filter on timestampColumnName. Omitting it is a full scan and often rejected by the downstream skill.
• Preserve workspaceId + itemId per binding — ontology bindings can reference source items in different workspaces from the ontology itself; do not assume collocation.
• Consult the in-memory grounding before issuing metadata calls — once the grounding JSON is decoded for the session, every subsequent walk / query reads source-column names, linking-table names, item GUIDs, clusterUri, and databaseName from grounding, not from a fresh list items / get eventhouse round trip. Re-fetching metadata you already have is the #1 cause of bloated call counts.
• Use GUIDs, not friendly names, in source URLs — many tenants run with FriendlyNameSupportDisabled, which silently rejects names like MyLakehouse.Lakehouse in OneLake DFS / Fabric REST URLs. Always pull the GUID from grounding and substitute it into the URL.

Prefer

• az rest with --body @file.json for any downstream KQL / SQL payload that contains |, ", or newlines. Inline --body breaks under shell escaping — see EVENTHOUSE-CONSUMPTION-CORE.md.
• Grounding summary JSON (see schema above) over raw definition.json dumps when handing context to another agent.
• take 100 / TOP 100 on first read of any entity's data, then refine.
• Inline az rest / curl over Python scripts for read-only consumption work — graph walks, single-entity lookups, and ad-hoc fan-out reads should compose ≤ ~15 REST calls in the shell. Reach for a script only when the work is stateful (envelope assembly, ID maps, LROs that the user wants re-runnable). See Snappy-Response Discipline and graph-walks.md.
• Cache the decoded definition for the life of one session — the ontology definition is orders of magnitude smaller than the source data and rarely changes mid-task. Refetch if the user mentions authoring activity.
• project / SELECT only the bound source columns, not the full physical table — ontology bindings imply an explicit column whitelist.

Avoid

• Querying the source table using ontology property names — those do not exist in the physical schema. Always go through propertyBindings[].
• Dropping where <timestampColumn> > ago(...) on TimeSeries / KustoTable reads — full scans on streaming tables are the #1 query failure.
• Joining across datasource kinds in one delegate call — if a relationship's two sides live in different source kinds, the delegate cannot express the join. Fetch both sides separately and join in the agent.
• Mutating the ontology from this skill — route all schema changes to fabriciq-ontology-authoring-cli.
• Passing raw base64 parts to downstream skills / models — always decode and reshape into a grounding JSON first.
• Silently ignoring unknown part paths — a new preview release may add part kinds; log-and-continue is fine but surface the new part name to the user.

Troubleshooting

Agentic Workflows

Snappy-Response Discipline

Consumption is read-only and should feel fast. The default mode is inline composition in the shell — not a Python wrapper. Use this checklist before reaching for a script:

Inline-first principles

1. Compose, don't script. Each REST call is independently meaningful; chain with && / | / jq rather than wrapping in a .py.
2. Parallelize fan-out. Independent reads (linking-table queries, per-entity-type IN-list reads) go in & with wait, not sequential for loops.
3. One round trip per group. WHERE id IN (...) not N×WHERE id = '...'.
4. Cache the grounding JSON for the session — it does not change between queries, and refetching it doubles every walk's latency.
5. Stream early. Show the anchor row and first neighbor group as soon as they return; don't block presentation on the full walk.
6. No script unless asked. If the user says "save this for later" or "re-run nightly", then package as a script. Otherwise keep the work in chat-replayable shell.

If a task starts to look like it needs > ~20 REST calls, > 2 hops, or persistent state across calls, surface this to the user before scripting — it usually means the question should be narrowed, not automated.

"Ground then Query" Sequence

When the user asks to query data through an ontology lens:

Step 1 → Resolve WS_ID + ONT_ID (list workspaces, list items type=Ontology)
Step 2 → getDefinition (LRO) → decode all parts → reconstruct tree
Step 3 → Build grounding JSON (entities, properties, bindings, relationships)
Step 4 → Disambiguate with the user if multiple entity types / bindings could satisfy the intent
Step 5 → For the chosen entity type + binding:
           a. Remap ontology property names → source column names
           b. Compose the source query (KQL / Spark SQL / T-SQL)
           c. Hand off to the matching sibling consumption skill with minimal fields
Step 6 → Post-process results back into ontology-property naming for the user (optional but helpful)

Schema-Aware Query Generation

After a grounding pass, generate queries using the physical columns recorded in propertyBindings[], never the ontology names:

Entity type "Aircraft" with binding:
  source kind = KustoTable
  sourceTableName = "AircraftReadings"
  timestampColumnName = "PreciseTimestamp"
  propertyBindings: { AltitudeFt → SourceColumn "Temp_C", TailNumber → SourceColumn "AssetId" }

User intent: "show altitude excursions on aircraft N42ZA in the last hour"

Generated KQL (delegated to eventhouse-consumption-cli):
  AircraftReadings
  | where AssetId == "N42ZA"
  | where PreciseTimestamp > ago(1h)
  | project PreciseTimestamp, Temp_C
  | where Temp_C > 80

Relationship Traversal

Relationship "operates" (Airline → Aircraft)
Contextualization: LakehouseTable "HubAircraftAssignment" with (AirlineId, TailNumber) columns
  AND two entity-type bindings (Airline on LakehouseTable "Airlines", Aircraft on KustoTable "AircraftReadings")

User intent: "which aircraft does Airline 'ZA' operate and what's their latest reading?"

Step 1: sqldw-consumption-cli → SELECT TailNumber FROM HubAircraftAssignment WHERE AirlineId = 'ZA'
Step 2: eventhouse-consumption-cli → AircraftReadings | where AssetId in (<TankIds>) | summarize arg_max(PreciseTimestamp, *) by AssetId
Step 3: Merge results in the agent; present with ontology-level column names.

Full end-to-end bash recipes (enumerate → ground → route → query) live in examples.md. For "show me everything related to X" prompts, jump straight to the N-hop walk in graph-walks.md.

Graph Walks (N-hop from an anchor)

When the user gives an instance ("Panel7", "aircraft N42ZA", "customer 1234") and asks for its neighborhood — not a single column — use the dedicated graph walk pattern instead of inventing a recipe per question.

Anchor → relationships touching anchor (from grounding JSON, in-memory)
       → linking-table reads (parallel, one per relationship)
       → IN-list reads of neighbor entities (one per EntityType)
       → optional KustoTable telemetry sweep (one per EntityType with TS binding)

For hop=1 this is typically ≤ 10 round trips and stays inline. Full algorithm, fan-out template, hop budget, and a worked Panel7 example live in graph-walks.md.

Examples

End-to-end worked examples (enumerate an ontology → build grounding JSON → route a source-table query to the correct sibling consumption skill → traverse a relationship across Lakehouse + Eventhouse) live in examples.md. N-hop neighborhood walks from an anchor instance (Panel7-style) live in graph-walks.md. Complete fetch-and-decode scripts live in grounding-extraction.md. Per-binding-type invocation templates live in routing.md.

Agent Integration Notes

• This skill is read-only on the ontology item. All authoring operations (create, alter, rebind, rename) belong to fabriciq-ontology-authoring-cli — delegate there.
• This skill does not execute data queries itself. It produces grounding context and a routing decision; the actual source query runs inside the per-datasource consumption skill you delegate to.
• Supported downstream skills (preview): eventhouse-consumption-cli, spark-consumption-cli, sqldw-consumption-cli. A standalone graph consumption skill does not yet exist — if the user asks for a graph-shaped query over relationships, surface that limitation and fall back to per-edge joins via Lakehouse / SQL.
• Orchestrator agents should hand this skill the workspace name / ID and the ontology display name / ID; it will return a grounding JSON they can reuse for subsequent delegate calls without re-fetching the definition.
• When authoring activity is suspected mid-session (e.g., the user runs fabriciq-ontology-authoring-cli and then comes back to query), re-run the grounding pass — the cached definition is stale.
• If a Fabric KQL MCP server is configured in the user's environment, it can substitute for az rest on the Eventhouse delegate leg. The repo's default mcp-setup/mcp-config-template.json does not register a fabric-kql server, so do not assume that name exists. Either way it does not cover ontology control-plane calls, so the fetch / grounding step still uses az rest.