How to play Duck Retriever
Systems that answer questions from documents (often called RAG, retrieval-augmented generation) first split each document into chunks, then search for the chunks that best match the question, and hand those to a language model to answer from. You play the search step.
Each question comes with four chunks from the same made-up document: a lease, a warranty, a policy, a manual. Pick the one that actually answers the question. Then every chunk shows why it does or doesn’t:
- Too broad: on the topic, but too general to answer.
- Wrong section: shares the question’s words, but it’s about something else.
- Missing the key number: the right subject, but not the figure the question asks for.
- Out of date: answered it once, but a newer version replaces it.
- Wrong case: a rule for a different situation from the one asked about.
Click a chunk or press 1 to 4. A practice round is eight questions; the daily round is five, the same for everyone.
Tips
- Find the exact thing the question asks for (a number, a day, a yes or no) and look for the chunk that states it.
- Watch for the situation: “at the end of the lease” and “leaving early” are different rules.
- Several chunks will share the question’s words. That’s exactly what fools keyword and vector search.
Questions
Are the documents real? No. Every document, rule and figure is made up for the game.
Where can I learn more? The free retrieval and RAG course covers chunking, embeddings, hybrid search, reranking and how to measure retrieval.
Is it free? Yes, with no account. Results are kept in this browser.