Advanced RAG Patterns Interview Guide: Agentic, Graph, and Corrective Retrieval

Advanced RAG pattern interviews separate architects from practitioners who only know naive vector search. Interviewers ask when agentic retrieval—query decomposition, iterative search, tool-augmented lookup—justifies extra latency and cost versus a single hybrid retrieval pass.

GraphRAG and knowledge-graph augmentation help multi-hop reasoning over entities but introduce graph maintenance overhead. Candidates explain build pipelines, community summarization costs, and when graph edges become stale faster than vector indexes can be refreshed.

Corrective RAG and self-RAG add critique loops that reject weak retrieval before generation. Discuss failure modes when critique models are miscalibrated, latency budgets for multi-pass pipelines, and eval proving that correction steps help on your query distribution—not just benchmarks.

Query transformation techniques—HyDE, multi-query expansion, step-back prompting—each shift recall-precision trade-offs. Interviewers want you to match technique to query type: acronym disambiguation versus comparative analytics versus procedural how-to questions.

Parent-child chunking, contextual retrieval, and late chunking address context loss at chunk boundaries. Articulate storage implications, reranking strategies, and how you measure whether advanced chunking fixes citations or only inflates index size.

The meta-skill is knowing when not to use advanced patterns. Production engineers propose baseline RAG with strong eval, then add complexity only when slice metrics prove bottlenecks. Interviewers reward disciplined iteration over pattern collecting.