Knowledge Extraction Produces Incorrect Relationship Assertions

Category: Document Ingestion / Knowledge Extraction Pipeline

Summary

A reproducible bug has been identified in the document ingestion and knowledge extraction pipeline. The extractor asserts a relationship that is directly contradicted by the source document it processed.

Bug 1 — Generation Collapse in Relationship Extraction

When a document establishes a three-step relationship chain (A is the parent of B, and B is the parent of C), the extractor collapses this to a direct two-step relationship (A is the parent of C), dropping the intermediate generation entirely.

The likely trigger is name co-occurrence — when A and C share a surname, a naming connection, or appear in close proximity in the text, the extractor treats proximity as a direct relationship rather than tracing the full chain stated in the document.

The result is a factually incorrect extracted node that directly contradicts the source document. In genealogical work, where the precision of the generational chain is the core purpose of the record, this is a significant accuracy failure.

Expected behavior: The extracted relationship should match what the source document explicitly states.

Actual behavior: The extractor asserts a direct parent-child relationship between A and C, skipping B entirely.

Suggested fix: Relationship extraction should require an explicitly stated or strongly implied direct link between two entities before asserting a parent-child relationship. Shared surnames or naming conventions alone should not be sufficient to assert direct parentage. Verifying the intermediate step before collapsing a chain would eliminate this error.