How do you debug a wrong RAG answer?
Start with the trace: question, retrieved chunks, ranks, scores, filters, context, citations, prompt, answer, and index version. Verify whether gold evidence was retrieved before blaming the model.
Interview Questions: Chunking Failure Analysis
Use these to explain chunking problems with production-level clarity.
What is lost context?
The retrieved chunk contains partial evidence but misses the surrounding definition, exception, warning, or condition needed for a faithful answer.
How do tiny chunks fail?
They match keywords but lack complete meaning. The model receives fragments instead of a usable thought unit.
How do giant chunks fail?
They retrieve broadly, consume tokens, weaken citations, and bury exact evidence inside unrelated context.
When is chunking not the root cause?
When gold evidence exists in the index but ranks too low, filters exclude it, reranking fails, prompt formatting hides it, or the model ignores good context.
How do you reduce overlap noise?
Lower overlap, deduplicate near-identical chunks after retrieval, rank diverse evidence, and monitor duplicate rate.
What metrics matter for chunking evals?
Recall@k, gold chunk rank, duplicate rate, stale-hit rate, boundary defect rate, parent-link completeness, citation support, and answer completeness.
How do you safely release a new chunking strategy?
Version the chunker and index, run golden queries, compare against baseline, inspect regressions, release behind a gate, and keep rollback ready.
How do you handle stale chunk failures?
Track active status, source version, effective date, and index release id. Filter old chunks and validate cleanup during re-indexing.
What is the best answer if an interviewer asks the ideal chunk size?
There is no universal size. The right size depends on corpus structure, question type, context budget, retrieval metrics, citation quality, and failure analysis.