Wikontic: ontology-driven knowledge graph extraction
It’s a graph-as-reasoning-substrate story with some interesting metrics. Most of all, it’s a lot cheaper than other graph RAG approaches.
It has a three-stages constraint Integration. It starts with candidate triplet extraction, followed by ontology-aware refinement, followed by entity normalization and alias-aware deduplication. Each stage tightens the graph a bit more. By stage 3 more than 96% of triplets are ontology-consistent. The remaining bits aren’t garbage but they’re flagged and interpretable. Rigor comes to mind.
Here’s the jaw-dropper: using the KG alone without touching source text Wikontic achieves 76 F1 on HotpotQA and 60 F1 on MuSiQue (competitive with text-augmented baselines). KG construction costs: 881 output tokens (Wikontic) vs 2K (AriGraph) vs 20K (GraphRAG). Efficiency by design.
The lesser. It’s Wikidata-only (even though the pipeline is domain agnostic). Every extraction stage uses LLM prompting and token costs are compute costs, not latency/throughput metrics. But the core insight holds: ontology-driven extraction, not post-extraction cleanup. Wikontic proves that knowledge augmented AI works.
Wikontic paper and code: https://yurakuratov.github.io/research/wikontic