GraphRAG Techniques: interview questions

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

1. Larkspur's cancellation answer reaches an export node but misses the retention condition. Where do you look first?

What a strong answer covers:

• Inspect the retrieved seed IDs, traversed relations and final passage IDs before editing the prompt.
• Check whether the retention relation was extracted, linked and reachable in the allowed direction.
• Separate an expansion cutoff from a reranking or context-budget omission.
• Add the complete policy path to the golden question's evidence set and check its sources.

2. Support asks for themes across the whole handbook, but the current retriever returns only cancellation neighbours. How would you change retrieval?

What a strong answer covers:

• Route the overview question to community reports with explicit corpus coverage.
• Check which communities and report levels contributed to the answer.
• Compare the overview against source passages and a flat baseline on the same corpus.
• Retain local retrieval for questions about a specific product or policy condition.

3. GraphRAG improves an offline answer score but indexing is too expensive after every handbook edit. What would you measure?

What a strong answer covers:

• Record extraction, embedding and report-generation work separately from query work.
• Measure update frequency and which reports depend on each changed source.
• Include failed calls, retries and storage in the comparison.
• Compare amortised work over the actual question volume without assuming an unmeasured break-even point.

4. Two documents use “Snapshot” for different products, and the assistant joins their policies. How do you contain the error?

What a strong answer covers:

• Inspect entity identity, product scope and version fields before increasing traversal depth.
• Split ambiguous entities and preserve the source record for each relation.
• Filter candidate passages and graph edges by the caller's product and access rights.
• Test a question where the wrong merge would produce a plausible answer with a false policy condition.

5. Every answer has numbered citations, yet reviewers keep finding unsupported claims. What would a useful attribution check do?

What a strong answer covers:

• Split the answer into claims and resolve each reference to an existing source span.
• Check that the span supports the claim and preserves its conditions and date.
• Audit relation evidence as well as the final node's description.
• Treat fixed confidence values and valid citation syntax as insufficient evidence of support.