Knowledge Graph Engineer Interview Questions for AI Training Work

I model at the level of granularity the actual queries need, rather than trying to capture every possible distinction upfront, since an overly granular model adds maintenance cost without a corresponding query benefit. I revisit the granularity if new query patterns emerge that the current model can't answer efficiently.

I use a combination of exact identifier matching where available and similarity-based matching on attributes like name and other distinguishing fields where it isn't, and I keep a mapping table auditable rather than silently merging entities. Incorrect merges are much harder to detect and undo than incorrect matches caught during review.

I favor a modular ontology structure where core concepts are stable and extensions can be added without restructuring the base, rather than trying to predict every future need upfront. Over-engineering the ontology for speculative future cases tends to add complexity that never gets used.

I filter as early as possible in the query pattern to reduce the intermediate result set before joining across more relationships, rather than joining broadly and filtering at the end. I also check the query plan or use profiling tools when a query is slower than expected, rather than guessing at the bottleneck.

I track provenance for each assertion so contradictions can be traced back to their source rather than silently overwritten, and I apply a defined precedence rule, like trusting a more authoritative source, rather than resolving conflicts arbitrarily. Losing the ability to trace where a fact came from makes future conflicts much harder to resolve.

I'd check whether the query relies on a pattern that scales poorly with graph size, like an unbounded traversal, and whether appropriate indexes still cover the query's access pattern. Graphs that grow substantially sometimes need index or schema adjustments that weren't necessary at a smaller scale.

I'd map the external ontology's concept to the closest equivalent in the existing schema rather than adopting it wholesale, documenting where the definitions diverge so downstream consumers understand the difference. Merging conflicting concepts as if they were identical creates subtle correctness issues that are hard to trace later.

I'd run consistency checks against the ontology's constraints and spot-check a sample of entities against source data directly, rather than assuming the ingestion pipeline produced correct output. Automated ingestion at scale tends to introduce errors that are only visible when you actually inspect specific records.

A graph modeled with a highly connected central entity became a bottleneck for nearly every query as usage grew, since almost every traversal passed through it. I restructured part of the schema to reduce that central dependency, which required migrating existing data but significantly improved query performance.

A stakeholder questioned why we needed a graph model instead of the relational database already in use. I showed a specific query, finding indirect relationships several hops away, that was straightforward in the graph model but required many complex joins in the relational approach, which made the tradeoff concrete.