Entity relationships (v1)

Entity extraction

Every article ingested by Atomize AI is scanned for named entities — companies, people, products, technologies, and concepts. Each detected entity is recorded in entity_mentions with a confidence score between 0 and 1 that reflects the extractor's certainty that the string refers to the canonical entity. Low-confidence mentions (below 0.5) are stored but excluded from graph edges and card counts by default.

Relationship inference

A relationship is recorded whenever two entities appear in the same article. Each co-occurrence is stored as a row in entity_pair_evidence and carries the article ID, the extracted relation_type (e.g., develops, used by, invests in), and a supporting text excerpt. Where the extractor cannot classify a relation type the row is tagged co-mentioned — still a signal, but a weaker one.

Confidence and weighting

Each relationship edge is weighted by three factors: mention frequency (how many articles link the pair), source diversity (how many distinct publishers contributed evidence), and evidence agreement (whether the extracted relation types are consistent across articles). Edges with higher frequency and source diversity rank higher in the Explorer and Explore views. Edges where evidence agreement is low are surfaced in the Diverge view as potential contradictions.

Signal, not consensus

The same disclaimer that appears on every Diverge card applies across the entire graph: a high-weight edge means the pair is frequently cited together across diverse sources — it does not mean the relationship is factually settled, widely agreed upon, or peer-reviewed. Relationship data reflects what publishers have written, not what is necessarily true.