Artefacts Composability Resolution (ACR)

Artefacts Composability Resolution (ACR) is a feature of the engine in OrionBelt Semantic Layer that answers a single, practical question while you build a query: given what you have selected so far, which other artefacts can you still add and get a valid result?

A semantic model is a graph of data objects connected by joins, with dimensions, measures, and metrics defined on top. Not every combination of these is valid: combining the wrong artefacts can multiply rows (fanout) or simply have no join path. ACR walks the model's join graph from your current selection (the anchor) and resolves the exact set of artefacts that remain composable with it:

• Dimensions you can group by, reachable through fanout-safe joins.
• Measures and metrics you can aggregate, drawn from the facts your selection belongs to.
• Cross-fact measures that are still combinable through the Composite Fact Layer (CFL), the UNION ALL planner for independent fact tables, surfaced separately so you know they join at a higher level.

The result is the composable set: a precise, deterministic list, not a guess. Because ACR is driven by the same join logic the compiler uses, anything it offers is guaranteed to compile.

ACR is an API capability first. The composables endpoint is the product: AI agents, BI tools, and custom applications call it to turn query construction into safe artefact composition, selecting from named artefacts known to combine instead of reasoning about table relationships and risking invalid SQL. The Gradio playground is just one consumer of that endpoint: it highlights composable artefacts in its pickers, but any frontend built on the API gets the same guarantees.

How it works

Join edges are directed from the source data object to its joinTo target, and a fact declares a many-to-one join to each of its dimension tables, so edges point fact -> dimension. ACR classifies artefacts by reachability from the query's grain:

Direction from the anchor	Yields	Why it is safe
Descendants (the dimension side)	dimensions to group by	reached via many-to-one joins, so no row multiplication
Ancestors (the fact side)	measures and metrics	the anchor is a dimension of those facts

An artefact is directly composable when it shares a single fanout-safe common root with the current anchor (a star query). A measure or metric whose fact is independent is CFL-composable when that fact still reaches the current grouping dimensions; it can then join as a separate UNION ALL leg. ACR reports the two groups separately.

The composables endpoint

Two ways to supply the anchor; both return the same response shape.

Query as the anchor (recommended)

Post the in-progress query. Its select.dimensions and select.measures (measures and metrics) form the anchor; the response lists what you can still add.

curl -X POST "$API/v1/sessions/$SID/models/$MID/composables" \
  -H "Content-Type: application/json" \
  -d '{ "select": { "dimensions": ["Customer Country"], "measures": ["Revenue"] } }'

{
  "anchorObjects": ["Customers", "Orders"],
  "dimensions": ["Customer Country", "Order Date", "Product Category", "Product Name", "Customer Segment"],
  "measures": ["Order Count", "Average Order Value", "Grand Total Revenue", "US Revenue"],
  "metrics": ["Revenue per Order", "Revenue Share"],
  "cflMeasures": [],
  "cflMetrics": []
}

dimensions / measures / metrics are directly composable. cflMeasures / cflMetrics are composable only via the Composite Fact Layer (an independent fact, unioned at the shared grain).

Named anchors

Pass one or more anchor names (a dimension, measure, metric, or data object). Repeated anchors are intersected:

curl "$API/v1/sessions/$SID/models/$MID/composables?anchor=Customer%20Country"

An empty anchor (an empty query, or no anchor parameter) means a fresh query: everything is composable.

Top-level shortcut

When exactly one model is loaded, the session/model path can be omitted:

curl "$API/v1/composables?anchor=Revenue"
curl -X POST "$API/v1/composables" -d '{ "select": { "measures": ["Revenue"] } }'

Relationship to other safety models

ACR is advisory. It tells you what composes and lets the compiler do the rest; it never blocks a selection. The UI highlights rather than hides, and cross-fact measures stay available as CFL candidates. That sets it apart from two other approaches semantic layers take:

• OLAP restriction (for example AtScale): valid combinations are implicit in the cube's conformed dimensions, and incompatible members are simply not offered.
• Compute-correct generation (for example Malloy): rather than guiding the selection, the compiler accepts the query and emits path-safe SQL using join-grain declarations (join_one / join_many) and symmetric aggregates, so fanout never corrupts a sum. OrionBelt instead keeps measures fanout-free by construction and unions independent facts through the CFL.

ACR is complementary to correctness, not a substitute: it shapes what a consumer composes, while the compiler still guarantees the SQL. See the comparison matrix for how each tool exposes (or does not expose) composability discovery.

In the UI

The Gradio playground marks composable artefacts in the Dimensions and Measures / Metrics pickers as you edit the query: a check mark for directly composable artefacts, and (via CFL) for cross-fact candidates. Highlighting never hides artefacts, so you can still start an independent (CFL) analysis at any time; the compiler remains the final validator.