Agentic Graph Repair

Knowledge graphs accumulate errors — missing relationships, constraint violations, outdated facts, and structural inconsistencies. Rather than relying on humans or hard-coded rules to fix them, we are building agentic pipelines where LLMs iteratively detect and repair graph inconsistencies in a closed loop.

The system is structured around four components: an Inspector that scans the graph and builds a repair queue, a Loader that retrieves focused subgraph context for each issue, an Analyst (LLM) that reasons about the inconsistency and generates a corrective Cypher query, and a Surgeon that applies the repair and logs the result. The loop continues until the repair queue is empty or an iteration limit is reached.

We handle two classes of inconsistency:

• Property-based: violations of attribute constraints (e.g., negative ages, date inversions)
• Label/structural: violations of schema rules (e.g., mutually exclusive labels, missing required relationships)

Early experiments used real-world datasets — a legal citation network and a protein-protein interaction network (STRING) — but both revealed a methodological challenge: without injected ground truth, LLM performance cannot be reliably measured. We pivoted to abstract synthetic graphs where anti-patterns are programmatically injected, enabling clean precision/recall evaluation.

This extends prior empirical work on single-turn LLM graph repair toward multi-step, tool-using agents capable of handling realistic repair scenarios at scale.

Team: Narendra Devireddy, Aadi Parakh, Tathagat Rath, Yash Vardhan Singh