Agentic Knowledge Graphs over PubMed
A citation and ontology knowledge graph built from structured PubMed metadata rather than LLM extraction, used to answer biomedical questions with traceable evidence and no answer when a claim cannot be traced.
This project builds a knowledge graph across the full PubMed 2026 baseline using citation and ontology metadata, avoiding LLM-based extraction entirely. On 600 held-out PubMedQA questions it reports 83.2% accuracy, compared with a 53.3% majority-class baseline and 37.2% for the same model given no evidence, with every answer traced back to a source path.
The graph contains more than 929 million edges built without any LLM calls, and the analysis is reproducible on a laptop using only Python's standard tooling plus numpy and matplotlib, with no GPU or corpus download required. The project also documents its own methodology under adversarial re-examination, reporting where earlier internal conclusions were later overturned by later measurement.