RAG Interview Questions
Dashboard
0%
1
Curious builder0 XP earned · 300 to level 2
0 daysFinish a lesson to begin
Badge collection0 of 6 unlocked
51 small wins to finish your pathNext question →

Q39HardConcept

What is GraphRAG? When would you use a knowledge graph with RAG?

30-second answerSay your answer out loud first, then reveal.
Documents go through LLM entity and relation extraction into a knowledge graph with community detection and summaries; a query is classed as local (entity lookup and neighbourhood traversal plus chunks) or global (map-reduce over community summaries), and both feed the LLM answer.

Where vector RAG fails and graphs help

  • Multi-hop: "Who manages the team that owns the payments service?" requires following edges (service → team → manager). The facts live in different documents that aren't semantically similar to the question.
  • Global questions: "What are the top recurring complaints this year?" can't be answered from the top 5 chunks.
  • Explainability: paths in the graph show why an answer was reached.

Costs and risks

  • LLM extraction over the entire corpus is expensive and slow.
  • Extraction errors (duplicate entities: "Krish", "Krish Naik", "K. Naik") need entity resolution.
  • Graph maintenance on updates is non-trivial.
  • Overkill for simple FAQ-style Q&A.

Lighter alternatives

  • Use an existing structured source (a CMDB, HR system, product catalogue) as the graph, with no extraction needed.
  • Iterative or agentic retrieval for multi-hop questions (Q35).
  • Metadata-rich chunks plus SQL for relational queries.
Interview framing. "I'd use GraphRAG when the questions are relational or global and the eval shows vector RAG failing on them, not by default."
PreviousHow would you scale RAG to 100 million+ documents (billions of chunks)?

This is what real progress feels like.