Executable Ontology Fixes Graph Retrieval Hallucinations

coldoven · reddit · 2026-07-13

The author tackles a classic graph retrieval issue: the graph contains an edge, but the edge type is incorrect, ultimately returning a highly confident but wrong answer.

Solution

Transforms the ontology from a document into executable rules, explicitly declaring relationship domain/range in YAML.
Validates edge legality on every traversal; if an invalid edge type is encountered, it throws a named error instead of silently returning the wrong node.

Additional Takeaways

The author discovered that the same declarations can be directly used for ranking:

Models the query as a small circuit, with the starting entity and target category acting as two poles.
Scores entities based on the current flowing through the nodes.
This filters out incorrect paths while naturally surfacing relevant entities at the top.

Example

For queries like "Sci-Fi by Nolan," the system correctly ranks true matches while assigning a score of 0 to nodes like The Dark Knight (which is related to Nolan but the wrong category) without needing extra manual filters.

The author notes this method is essentially based on classic harmonic functions / current-flow centrality, with the novelty being the unification of these concepts with knowledge graph validity checks from a single source. The project offers an API compatible with 9 graph database backends and can run offline.

Original post →

Executable Ontology Fixes Graph Retrieval Hallucinations

More from coding & agent