Knowledge Graph (retrieval)

The Knowledge Graph inside Concord by IaxaI indexes resolved entities, normalized events, and extracted (subject, relation, object) triples into a dense-sparse graph. Phrase nodes hold entities, controls, regulations, techniques, and indicators. Passage nodes hold events, documents, alerts, and incidents. Edges connect them through three relationships: relation, synonym, and context. Retrieval is a fixed pipeline. The query embeds with a 768-dim sentence transformer, dual-path search hits both phrase and passage indexes, threshold filtering trims the candidates, Personalized PageRank seeds traversal from the surviving nodes, and the final score blends vector similarity, PPR weight, and recency. Three latency tiers cover lookup under 100ms, semantic retrieval under 500ms, and deep analysis under 5s. No language model runs in the query path. LLM-assisted information extraction runs only in nightly batch enrichment. The knowledge graph is what every surface (Dedup, Detection Portability, Compliance Auto-Packets) queries when it needs context beyond a single event.