AIML Rounds

QUIZ

RAG

12 questions on RAG, from 12 lessons. Pick an answer in your head, then open it.

On the roadmap for: AI Engineering, ML Engineering.

  1. Question 1 of 12

    What are the basic steps of a RAG system?

    1. A. Put the full text of every document into each prompt
    2. B. Chunk and index documents, retrieve the best chunks, answer from them
    3. C. Fine-tune the model on every document so it memorises them
    4. D. Search the public web for each question, then summarise the top results for it
    View answer

    ✓ B. Chunk and index documents, retrieve the best chunks, answer from them

    RAG keeps knowledge outside the model and fetches it at question time, so it can be updated without retraining.

    Lesson: RAG End to End

  2. Question 2 of 12

    Chunks that are too large cause one problem; chunks that are too small cause another. Which pair is right?

    1. A. Large: slower to embed and store; small: slower to retrieve
    2. B. Size only affects storage cost, not the quality of answers
    3. C. Large: answers get diluted; small: answers lose context
    4. D. Large: always more accurate; small: never retrieved at all
    View answer

    ✓ C. Large: answers get diluted; small: answers lose context

    Chunk size is a trade-off between precise retrieval and enough context to answer; overlap and structure-aware splitting help.

    Lesson: Parsing and Chunking

  3. Question 3 of 12

    Your RAG system misreads tables in PDFs. Where should you fix it first?

    1. A. In parsing: extract tables with rows and headers intact
    2. B. Switch to a bigger LLM that reads tables better
    3. C. Retrieve more chunks per question to cover the whole table
    4. D. Raise the temperature so the model explores more
    View answer

    ✓ A. In parsing: extract tables with rows and headers intact

    If parsing flattens a table into jumbled text, no retriever or model can recover it. Garbage in, garbage out.

    Lesson: Parsing and Chunking

  4. Question 4 of 12

    Approximate nearest-neighbour indexes such as HNSW trade what for speed?

    1. A. Nothing: they return exactly the true nearest neighbours
    2. B. They only work for text embeddings, not images
    3. C. A small loss in recall: some true neighbours may be missed
    4. D. They need labelled query–neighbour pairs before they can build the index
    View answer

    ✓ C. A small loss in recall: some true neighbours may be missed

    Exact search over millions of vectors is too slow, so ANN indexes search a graph or clusters and accept slightly imperfect recall.

    Lesson: Embeddings and Vector Indexes

  5. Question 5 of 12

    Why does pure dense (embedding) search often miss queries with exact product codes or names?

    1. A. Embeddings blur rare exact tokens; keyword search matches them
    2. B. Embeddings can't store numbers or symbols at all
    3. C. Tokenisers delete product codes and other symbols before the text is embedded
    4. D. BM25 is always more accurate than embedding search
    View answer

    ✓ A. Embeddings blur rare exact tokens; keyword search matches them

    Hybrid search combines both and merges the lists, for example with reciprocal rank fusion.

    Lesson: Advanced RAG: Hybrid Search, Re-ranking, Query Rewriting and GraphRAG

  6. Question 6 of 12

    The right chunk is retrieved but ranked 12th, and you pass the top 5 to the LLM. What's the usual fix?

    1. A. Add a re-ranker, such as a cross-encoder, on the top results
    2. B. Shorten the question so retrieval focuses better
    3. C. Pass all 100 retrieved chunks to the LLM
    4. D. Lower the temperature, so the model reads the retrieved chunks more carefully
    View answer

    ✓ A. Add a re-ranker, such as a cross-encoder, on the top results

    Retrieve wide and cheap, then re-rank a short list with a slower, more accurate model.

    Lesson: Re-ranking and Late Interaction

  7. Question 7 of 12

    What is 'late interaction', as in ColBERT?

    1. A. Calling the LLM again after the user has left the chat
    2. B. Delaying indexing until traffic is low at night
    3. C. Re-ranking the results with a second, larger LLM afterwards
    4. D. One embedding per token, scored by matching token vectors
    View answer

    ✓ D. One embedding per token, scored by matching token vectors

    It sits between a single-vector bi-encoder (fast, coarse) and a cross-encoder (slow, precise).

    Lesson: Re-ranking and Late Interaction

  8. Question 8 of 12

    The correct chunk was retrieved, but the answer is still wrong. Which is a likely cause?

    1. A. The question was too short for the model to understand
    2. B. The embedding model is too small to match the question to the right chunk
    3. C. The vector index was down, so nothing was really retrieved
    4. D. The model ignored or misread it, e.g. among many chunks or conflicts
    View answer

    ✓ D. The model ignored or misread it, e.g. among many chunks or conflicts

    Retrieval isn't the whole story: ask for answers grounded in quoted sources, handle conflicts, and allow 'I don't know'.

    Lesson: Grounded Answers: Citations, Conflicts and Saying No

  9. Question 9 of 12

    When might you skip building a vector index altogether?

    1. A. When users ask many questions every single day
    2. B. When the data fits in context or a tool can fetch it
    3. C. When the documents change several times a day
    4. D. Never: every LLM app needs its own vector index
    View answer

    ✓ B. When the data fits in context or a tool can fetch it

    Long context, structured queries or an agent with search tools can beat a vector index for small or structured data.

    Lesson: Agentic RAG, and When Not to Index

  10. Question 10 of 12

    How should a RAG system stop users from retrieving documents they're not allowed to see?

    1. A. Remove sensitive words from documents before indexing
    2. B. Tell the LLM in the system prompt not to reveal them
    3. C. Hide restricted documents in a separate, secret index only
    4. D. Store access rights with each chunk and filter at retrieval
    View answer

    ✓ D. Store access rights with each chunk and filter at retrieval

    Instructions can be ignored or bypassed; access filtering must happen before the model sees anything.

    Lesson: Permissions and Poisoning in RAG

  11. Question 11 of 12

    A better embedding model comes out. What must you do to use it?

    1. A. Re-embed the whole corpus, since the vector spaces differ
    2. B. Embed only new documents with the new model
    3. C. Nothing: vectors from any model can be compared directly
    4. D. Change the chunk size to match the new model
    View answer

    ✓ A. Re-embed the whole corpus, since the vector spaces differ

    Each model has its own vector space. Plan a background re-index and switch over once it's complete and evaluated.

    Lesson: Keeping an Index Current

  12. Question 12 of 12

    Your RAG chatbot gave a confident wrong answer. What should you check first?

    1. A. The trace: was the right content retrieved, and was it used?
    2. B. The number of users, since load causes wrong answers
    3. C. The model's size, since bigger models tend to hallucinate much less often
    4. D. The temperature setting, since it was probably too high
    View answer

    ✓ A. The trace: was the right content retrieved, and was it used?

    Separating retrieval failures from generation failures tells you which half of the system to fix.

    Lesson: RAG: Wrong Answer, Where to Look First