Knowledge Graphs with Neo4j: interview questions

Production questions of the kind asked in applied AI engineering interviews.

1. Larkspur's EU export plan disappears after a handbook update, although its passage still exists. How do you locate the failure?

What a strong answer covers:

• Checks ingestion version, entity key and the required feature and region paths.
• Separates a missing graph edge from a candidate that fell below the vector budget.
• Reads the extraction input and supporting passage before changing the prompt.
• Adds the updated plan and previous version to the fixed regression questions.

2. Product wants both a free-text search and exact regional availability. Would you retrieve vectors first or match the graph first?

What a strong answer covers:

• Uses exact graph predicates when identity and constraints are known.
• Uses semantic candidates for approximate descriptions and validates their graph relationships.
• Measures eligible-answer recall against candidate budget, query latency and graph degree.
• Applies authorisation across every retrieval branch and checks all returned source passages.

3. The assistant returns a convincing answer from an empty Cypher result. What should monitoring record?

What a strong answer covers:

• Logs original question, rewritten question, generated query, parameters and rows separately.
• Grades query results against known paths as well as grading final answer support.
• Tracks empty-result and truncation rates by question type and schema version.
• Requires an explicit unsupported-answer response when the available facts cannot answer.

4. A CSV re-import doubles the number of plans after descriptions change. How do you repair the import design?

What a strong answer covers:

• Inspects which mutable properties were included in the merge pattern.
• Uses a stable scoped identifier with an appropriate uniqueness constraint.
• Updates descriptions separately and resolves existing duplicates with source-aware review.
• Tests repeated imports and edited descriptions on a fixture before loading the full catalogue.

5. A generated query asks for a plan relationship that is absent from the schema. Should the system retry until it finds an answer?

What a strong answer covers:

• Checks whether the requested relation is representable and present in the current schema.
• Bounds retries and query execution, with read-only database permissions.
• Distinguishes a translation error from missing source data and ambiguous identity.
• Returns the evidence limitation rather than inventing a relation to satisfy the request.