RAG Prompt Engineering: Grounding, Citations, and Retrieved Context

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-prompting-prompt-retrieval-augmented-creators

1. Which instruction should a well-designed RAG prompt include to enable verification of responses?

   1. Tell the model to answer based on its pre-training knowledge
   2. Ask the model to generate a confidence score for each answer
   3. Request the model to summarize the retrieved information in bullet points
   4. Instruct the model to cite claims using chunk IDs from the retrieved context

2. A RAG system retrieves documents about car repairs, but the user's question is about flight booking. What should the prompt instruct the model to do?

   1. Use general knowledge about travel to answer anyway
   2. Search the web for additional flight booking information
   3. State that the retrieved context doesn't contain enough information to answer
   4. Combine the unrelated car repair facts into a response

3. Which limitation can a carefully crafted prompt NOT overcome in a RAG system?

   1. The model failing to cite any sources at all
   2. The retrieval system returning irrelevant or incorrect chunks
   3. The model making up citation IDs that don't exist
   4. The model refusing to use the provided context

4. What does 'grounding' mean in the context of RAG systems?

   1. Training the retrieval system on more diverse data
   2. Basing all responses strictly on facts present in the retrieved context
   3. Limiting the length of model outputs
   4. Ensuring the model generates grammatically correct sentences

5. A student notices that a RAG system's responses sometimes include citation IDs like [chunk_99] even when only 50 chunks were retrieved. What is happening?

   1. The retrieval system is malfunctioning and skipping numbers
   2. The user has asked too many questions, causing chunk ID overflow
   3. The chunks were retrieved from different documents with conflicting numbering schemes
   4. The model is generating fictional citation IDs—a known risk that requires programmatic verification

6. What does the lesson recommend when the retrieved context contains multiple chunks that partially address different aspects of a question?

   1. Use facts from all relevant chunks and cite each one appropriately
   2. Combine the chunks into a single paraphrase without citations
   3. Answer only from the first chunk to avoid confusion
   4. Ignore all chunks and answer from memory to ensure completeness

7. Why is programmatic verification of citations important even when using a well-designed RAG prompt?

   1. Because models may generate hallucinated or incorrect citation IDs
   2. Because users prefer seeing numeric IDs over text citations
   3. Because the prompt will eventually expire and need replacement
   4. Because the API will automatically validate citations anyway

8. A developer writes a prompt that says 'Use the context below to answer the question.' The RAG system still produces incorrect answers. What is likely missing from the prompt?

   1. An instruction to answer the question before reading the context
   2. A request for the model to be more creative
   3. Explicit instructions about how to handle irrelevant chunks and requirement to cite sources
   4. A warning not to use any numbers in responses

9. Which statement about chunk IDs in RAG systems is correct?

   1. Chunk IDs are automatically generated by the retrieval system and never need verification
   2. Chunk IDs are optional in RAG prompts since users can read the context themselves
   3. Chunk IDs should be included in the prompt as examples for the model to follow
   4. The model always generates accurate chunk IDs without instruction

10. A RAG prompt includes: 'Use ONLY facts from the context. Cite each claim.' What additional instruction would make the prompt more complete?

    1. Instruct the model to prefer its own knowledge over context
    2. Request a longer response with more detail
    3. Ask the model to add a creative story element to engage readers
    4. Tell the model what to do when context doesn't contain the answer

11. What distinguishes a strong RAG prompt from a weak one?

    1. Strong prompts ask the model to verify its own answers before responding
    2. Strong prompts are written by domain experts in technical language
    3. Strong prompts explicitly address source handling, citation, and handling of gaps in information
    4. Strong prompts are longer and include more examples

12. When a user asks about a topic and the RAG system retrieves zero relevant chunks, what is the ideal model behavior?

    1. State that there isn't enough information in the provided context to answer
    2. Refuse to respond and ask the user to rephrase the question
    3. Provide a guess based on the most similar retrieved document
    4. Apologize and suggest the user try a different search engine

13. A developer tests a RAG system and finds the model frequently cites [chunk_5] for facts that appear nowhere in chunk 5. What should the developer do?

    1. Remove all citation requirements from the prompt
    2. Increase the temperature parameter to reduce repetition
    3. Reduce the number of retrieved chunks to minimize confusion
    4. Implement programmatic verification to catch incorrect citations

14. What does it mean to 'bound output to facts present in chunks'?

    1. Ensuring the model only uses complete sentences from chunks
    2. Restricting the response to only facts explicitly stated in the retrieved context
    3. Limiting how many words the model can generate
    4. Requiring the model to cite at least three chunks in every response

15. Why might a RAG prompt that works well for one use case fail in another?

    1. Different use cases involve different types of retrieved context and different requirements for handling irrelevant information
    2. The model behaves differently on weekdays versus weekends
    3. Prompts are universally applicable and shouldn't need adjustment
    4. RAG systems require complete redesign when changing topics