Skip to main content

RAG Interview Questions

1. Corrective RAG (CRAG) vs. Self-RAG​

Question: How does a Corrective RAG (CRAG) strategy improve reliability compared to a standard pipeline when the retriever returns low-quality or irrelevant documents?

Answer: CRAG introduces a lightweight retrieval evaluator (usually a T5 model or a small LLM) that scores the relevance of each retrieved chunk.

  • If relevant: The chunks are passed to the generator.

  • If irrelevant: The system triggers a web search fallback or an alternative knowledge base search.

  • If ambiguous: It performs a mix of internal context and external search.

    Unlike standard RAG, which forces the LLM to "make do" with whatever it gets, CRAG acts as a quality gate, significantly reducing hallucination caused by "noise" in the retrieval.

2. Multi-Hop Reasoning with GraphRAG​

Question: Why does GraphRAG (Knowledge Graph + Vector Search) outperform pure Vector RAG for "global" queries like "What are the recurring themes across all 500 research papers?"

Answer: Vector search is inherently local; it finds specific snippets that are semantically similar to the query. It cannot "connect the dots" across an entire corpus. GraphRAG extracts entities and relationships into a Knowledge Graph and pre-summarizes "communities" (clusters) of nodes.

  • For local queries, it uses graph traversal to find related entities.

  • For global queries, it aggregates those pre-generated community summaries.

    This allows the model to answer questions that require a high-level synthesis of the entire dataset rather than just a specific page.

3. Handling "Lost in the Middle" in Long Context​

Question: With context windows now reaching 1M+ tokens, is RAG still necessary? How do you solve the "Lost in the Middle" problem in long-context models?

Answer: Yes, RAG remains critical for cost (1M tokens per query is expensive) and latency. Even with large windows, LLMs struggle to recall information buried in the middle of a massive prompt.

Advanced solutions include:

  1. Contextual Compression: Using a model to summarize retrieved chunks specifically in relation to the query before stuffing them into the prompt.

  2. Long-Context Re-ranking: Re-ordering the top-K results so that the most relevant information is at the extreme beginning or end of the prompt, where attention heads are most effective.

4. Agentic RAG and "Plan-and-Execute"​

Question: How do you implement a Plan-and-Execute agent for RAG, and how does it handle multi-step queries like "Compare the Q3 revenue of Company A and Company B, then explain the trend"?

Answer: A single retrieval step fails here because the model doesn't know the revenue values yet. A Plan-and-Execute agent:

  1. Planner: Breaks the query into sub-tasks: (a) Retrieve Co-A Q3 revenue, (b) Retrieve Co-B Q3 revenue, (c) Compare and analyze.

  2. Executor: A worker agent calls the RAG tool specifically for each sub-task.

  3. Re-planner: After each step, the agent checks if it has enough info.

    This Agentic Loop allows for "multi-hop" reasoning where the output of one retrieval informs the next query.

5. Evaluation with RAGAS "Faithfulness" vs. "Relevance"​

Question: In the RAGAS framework, explain the mathematical or logical difference between Faithfulness and Answer Relevance.

Answer:

  • Faithfulness (Groundedness): Measures if the answer is derived only from the retrieved context.

    • Method: Extract claims from the answer $\rightarrow$ verify each claim against the context using an LLM.

    • Goal: Stop the model from using its "internal knowledge" to fill gaps.

  • Answer Relevance: Measures if the answer actually addresses the user's question, regardless of whether it's true.

    • Method: The LLM generates 3 potential questions that the answer could be responding to, then calculates the cosine similarity between those and the actual user query.