Module 5: Knowledge Check — RAG & Data Grounding
Test your understanding of Retrieval-Augmented Generation and data grounding concepts.
Question 1
What is the primary purpose of RAG?
- A) To fine-tune a model on your data
- B) To ground LLM responses in retrieved, factual data
- C) To replace the need for prompt engineering
- D) To train a model from scratch
Answer
B — RAG retrieves relevant context at query time and provides it to the LLM, grounding the response in your actual data rather than relying solely on the model's training data.
Question 2
A user reports that your RAG chatbot often returns answers that don’t match the retrieved documents. Which metric best measures this problem?
- A) Fluency
- B) Coherence
- C) Groundedness
- D) Similarity
Answer
C — Groundedness measures whether the model's response is supported by the retrieved context. Low groundedness means the model is generating information not found in the source documents.
Question 3
Your documents contain complex technical manuals with headings, tables, and diagrams. Which chunking strategy is most appropriate?
- A) Fixed-size chunking with 256 tokens
- B) Sentence-level splitting
- C) Semantic chunking
- D) No chunking (full document)
Answer
C — Semantic chunking preserves logical structure by splitting at natural boundaries (sections, headings), which is critical for complex documents where context within a section matters.
Question 4
Which combination of search methods provides the best retrieval quality for production RAG?
- A) Keyword search only
- B) Vector search only
- C) Keyword + vector (hybrid) search
- D) Keyword + vector + semantic reranking (hybrid + reranking)
Answer
D — Hybrid search with semantic reranking consistently outperforms any single method. Keywords catch exact matches, vectors capture semantic meaning, and reranking uses deep learning to refine the ordering.
Question 5
You need to index 50,000 documents and want Azure AI Search to handle both chunking and embedding automatically. Which feature should you use?
- A) Custom skillsets
- B) Integrated vectorization
- C) Manual embedding pipeline
- D) Search index projections
Answer
B — Integrated vectorization automates both chunking and embedding during indexing, and provides query-time vectorization so you can send plain text queries.
Question 6
In a multi-turn conversation, a user says “Which one is cheapest?” after discussing Azure OpenAI models. What technique should you apply before searching?
- A) Increase the number of search results
- B) Query rewriting using chat history
- C) Switch to a larger embedding model
- D) Add more documents to the index
Answer
B — Query rewriting reformulates the ambiguous follow-up question into a standalone query using the conversation history, ensuring the search retrieves relevant results.
Question 7
Your enterprise RAG application must ensure that finance documents are only visible to the finance team. How should you implement this?
- A) Create separate indexes for each team
- B) Use document-level security filters with OData expressions
- C) Remove sensitive documents from the index
- D) Add instructions to the system prompt
Answer
B — Document-level security filters use OData expressions to restrict search results based on user group membership, ensuring each user only retrieves documents they're authorized to see.
Question 8
Which embedding model offers the highest accuracy for vector search in Azure OpenAI?
- A) text-embedding-ada-002
- B) text-embedding-3-small
- C) text-embedding-3-large
- D) gpt-4o
Answer
C — text-embedding-3-large produces 3072-dimensional vectors and provides the highest embedding accuracy among Azure OpenAI embedding models.
Question 9
Your RAG system retrieves relevant documents but the LLM’s answers are often incomplete. What should you investigate first?
- A) Switch to a different embedding model
- B) Increase chunk overlap
- C) Check if the system prompt instructs the model to use all retrieved context
- D) Reduce the number of retrieved documents
Answer
C — If retrieved documents are relevant but answers are incomplete, the issue is likely in the generation step. Review the system prompt to ensure it instructs the model to synthesize a comprehensive answer from all retrieved context.
Question 10
What is the recommended chunk overlap percentage to prevent losing context at chunk boundaries?
- A) 0% (no overlap)
- B) 5%
- C) 10–20%
- D) 50%
Answer
C — 10–20% overlap ensures that important context at chunk boundaries is captured in adjacent chunks, preventing information loss during retrieval.
Question 11
In agentic RAG, what distinguishes it from standard RAG?
- A) It uses a different embedding model
- B) The agent decides when and what to retrieve rather than always retrieving
- C) It doesn’t use a search index
- D) It requires fine-tuning
Answer
B — In agentic RAG, the AI agent evaluates each query and decides whether retrieval is needed, what to search for, and which data sources to query — rather than blindly retrieving on every request.
Question 12
You’re building a RAG application that searches product docs, support tickets, and an internal wiki simultaneously. What pattern is this?
- A) Conversational RAG
- B) Agentic RAG
- C) Multi-index RAG
- D) Cached RAG
Answer
C — Multi-index RAG searches across multiple data sources in parallel and merges/reranks the results to provide comprehensive answers from diverse knowledge bases.
Question 13
Your RAG system’s retrieval precision is 0.4 (only 40% of retrieved documents are relevant). Which improvement would have the most impact?
- A) Add a caching layer
- B) Increase the number of retrieved documents from 5 to 20
- C) Improve index design with better chunking, metadata filtering, and semantic reranking
- D) Switch from GPT-4o to GPT-4.1
Answer
C — Low retrieval precision means the search isn't finding the right documents. Better chunking, metadata filters, and semantic reranking directly improve retrieval quality. Increasing results would lower precision further.
Question 14
What is the benefit of using dimension reduction with text-embedding-3-large?
- A) It improves embedding accuracy
- B) It reduces storage costs while maintaining reasonable accuracy
- C) It increases the maximum token input
- D) It enables semantic reranking
Answer
B — Dimension reduction (e.g., from 3072 to 1024) reduces the storage required for each vector, lowering index costs while preserving most of the embedding's semantic quality.
Question 15
Before deploying a RAG system to production, you should evaluate quality. Which set of metrics is most relevant?
- A) Training loss, validation loss, perplexity
- B) Groundedness, relevance, faithfulness, retrieval precision
- C) Tokens per second, latency, throughput
- D) Model size, parameter count, context window
Answer
B — RAG-specific metrics like groundedness (is the answer supported by context?), relevance (are retrieved docs relevant?), faithfulness (does the answer accurately represent sources?), and retrieval precision (fraction of relevant results) directly measure RAG quality.
Scoring
| Score |
Rating |
| 13–15 correct |
⭐ Excellent — ready for production RAG |
| 10–12 correct |
👍 Good — review advanced patterns |
| 7–9 correct |
📖 Review lessons 1–3 |
| Below 7 |
🔄 Revisit the module from the beginning |