KG-First, LLM-Fallback: A Hybrid Microservice for Grounded Skill Search and Explanation
Hybrid system achieves 0.94 nDCG@5 in under 200ms latency.
A new paper from Ngoc Luyen Le, Marie-Hélène Abel, and Bertrand Laforge presents SkillGraph-Service, a hybrid microservice designed to bridge the gap between complex competency frameworks and practical educator use. The system unifies authoritative taxonomies like ESCO, ROME, and O*NET into a provenance-preserving Knowledge Graph (KG) while adopting a KG-first, LLM-fallback architecture that balances symbolic rigor with sub-symbolic flexibility.
Empirical evaluation on a multilingual dataset shows that the hybrid retrieval strategy (fusing SQLite FTS5 full-text search with HNSW vector search) achieves impressive accuracy—nDCG@5 over 0.94—with sub-200ms latency, suggesting expensive cross-encoder re-ranking is unnecessary in this domain. The authors also analyze LLM-generated explanations, finding that JSON-constrained outputs ensure high citation precision but that deterministic templates remain best for maximizing evidence coverage. This architecture offers a scalable, auditable solution for integrating skill data into digital learning ecosystems.
- SkillGraph-Service unifies ESCO, ROME, and O*NET into a single Knowledge Graph preserving provenance.
- Hybrid retrieval (SQLite FTS5 + HNSW) achieves nDCG@5>0.94 with sub-200ms latency.
- JSON-constrained LLMs provide high citation precision; deterministic templates maximize evidence coverage.
Why It Matters
Scalable, auditable skill search helps educators align curricula with real labor market demands.