scene_service.scene_graph.relations¶

Edge candidate generation and LLM-based relation inference.

Functions

`compute_geometry_hint`(a, b)
`generate_edge_candidates`(nodes, *[, ...])	Return candidate pairs with precomputed geometry hints.

Classes

Infer the spatial relation between two objects via LLM.

class scene_service.scene_graph.relations.RelationInferer(llm_client: SceneGraphLLMClient)[source]¶

Bases: object

Infer the spatial relation between two objects via LLM.

Call LLM to infer the relation from source to target.

Returns a SceneGraphEdge. On LLM failure the edge has relation="unknown" and method="llm_fail".

async infer_semantic_relation(source: SceneGraphNode, target: SceneGraphNode, hint: GeometryHint, known_relation: str) → SceneGraphEdge[source]¶: Semantic-only inference for a pair whose spatial relation geometry already owns (known_relation). The LLM is constrained to the semantic vocabulary; any spatial or off-vocabulary answer collapses to "none" (geometry stays authoritative for the spatial slot). On transport failure the edge is relation="unknown" / method="llm_fail" so the builder retries it like a rate-limited pair rather than caching a false “no semantic relation”.

scene_service.scene_graph.relations.compute_geometry_hint(a: SceneGraphNode, b: SceneGraphNode) → GeometryHint[source]¶

scene_service.scene_graph.relations.generate_edge_candidates(nodes: list[SceneGraphNode], *, max_distance: float = 2.0, min_xy_overlap: float = 0.15, max_candidates: int = 200) → list[tuple[SceneGraphNode, SceneGraphNode, GeometryHint]][source]¶

Return candidate pairs with precomputed geometry hints.

A pair is a candidate if any of these hold: