Expand description
Explainability engine for graph-based RAG — attention weights, path explanation, provenance.
Provides human-interpretable explanations of why a set of triples was retrieved for a given query. Three explanation strategies:
- Attention weights — normalized relevance scores per triple
- Path explanation — BFS shortest-path from query entity to result entity
- Provenance — source documents / IRIs that contributed to the answer
Structs§
- Attention
Explanation - Normalized attention weights over a set of triples.
- Explainability
Engine - Top-level explainability engine — wraps all explanation strategies.
- Path
Explanation - A BFS shortest-path explanation from a query entity to a result entity.
- PathHop
- One hop in a graph path.
- Provenance
Report - Source provenance for a set of triples.
- Scored
Triple - A knowledge graph triple with an associated relevance score.