Hierarchical & Memory-Augmented Retrieval: interview questions

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

1. Larkspur's assistant says cancelled accounts lose data immediately, although the source passage describes a retention window. The retrieved summary omits that window. How do you debug it?

What a strong answer covers:

• Trace the summary to the exact source version and inspect whether summarisation dropped the exception
• Inspect both the summary rank and the source passage rank before changing the answering prompt
• Retrieve source evidence for precise policy claims and flag unsupported summary assertions
• Add the cancellation question to the golden set and rebuild affected summary ancestors

2. Operations asks for themes across the handbook, while support asks exact policy questions. Would you force both through top-down tree traversal?

What a strong answer covers:

• Compare traversal with collapsed search on both question types under the same context budget
• Check whether choosing one high-level parent excludes a needed leaf early
• Retain precise passages when summaries lack the required condition or number
• Report indexing work and per-query quality separately instead of assuming a tree is faster

3. A new memory-assisted retriever returns fluent clues about a service Larkspur no longer sells. What should you monitor?

What a strong answer covers:

• Log memory version, generated clues, retrieved source identifiers and final citations
• Measure evidence recall and answer support separately from clue fluency
• Invalidate derived memory when its source documents change or are deleted
• Compare against original-query retrieval and fall back when clues find no supported evidence

4. Larkspur has two products with the same short name. Entity extraction merges them and connects the wrong cancellation rule. How should the graph change?

What a strong answer covers:

• Use stable identifiers and product context instead of case-normalised names alone
• Keep relation direction, qualifiers and source identifiers when the domain requires them
• Inspect extracted assertions against passages before accepting a merge
• Rebuild affected profiles and test an ambiguous-name query after the correction

5. A private incident report is deleted, but a global handbook summary still mentions it. What must deletion remove?

What a strong answer covers:

• Follow dependency links through summaries, memory entries, graph descriptions and vector indexes
• Remove or regenerate derived text that contains the report's information
• Apply access controls before derived evidence is exposed, including shared summaries
• Verify with a deletion regression query and inspect stored representations as well as the answer