Retrieval: hybrid, rerank, query transformation

8 questions on this module's lessons. Pass mark 70 %. Your result feeds interview readiness, and a failed checkpoint puts the module in your revision queue.

1. 1. A support bot's top hit for a refund question is about refunds but never states the refund window. When you print the retrieved chunks, you see the sentence giving the window landed in the next chunk, and that chunk was not retrieved. What is the right fix?

   Add a cross-encoder reranker over the top 20 so the right chunk rises to the top.Tighten the generation prompt so the model makes better use of the context it was given.Fix the chunking so the window sentence stays with its topic, then re-measure recall@5 on the failing questions.Raise the similarity threshold so only the strongest matches reach the prompt.
2. 2. Dense search ranks document A first and document B second. BM25 ranks document C first and document B second, and A does not appear in the BM25 list. Using reciprocal rank fusion with the conventional k of 60, which document comes out on top, and why?

   B, because it collects credit from both lists, and at k=60 a first place earns barely more than a second, so agreement outweighs one top spot.A, because the dense retriever's first place carries the highest similarity score.C, because a rare-token match in BM25 is stronger evidence than a semantic match.It cannot be decided until both score sets are normalised to 0 to 1.
3. 3. A help-centre corpus changes rarely and gets a very high volume of queries. Short user questions keep missing the long, answer-shaped chunks that hold the answers. Which technique fits best?

   HyDE, because writing a hypothetical answer for each query adapts at once to a changed corpus.Multi-query rewriting, because the user's words differ from the document's.Self-query retrieval, because the questions need structured metadata filters.HyPE, because generating the questions each chunk answers at ingestion removes the model cost per query.
4. 4. You turn on MMR with k=4 and turn λ down to favour diversity, but the results still match plain similarity search, near-duplicates included. What is the most likely cause?

   λ is still too high; set it to 0 so diversity is the only criterion.fetch_k is set to about the same value as k, so there is no wider pool of alternatives to choose from.MMR cannot act on chunks that overlap heavily, so the near-duplicates must be removed at indexing.k is too small; MMR needs a k of around 20 before it can diversify.
5. 5. A support search covers a help-centre wiki, tickets and release notes. Answers from the existing documents are reliable, but new error codes appear before anyone documents them. The data is internal and must not be sent to a public search. Which control loop and fallback fit?

   Self-RAG, because grading the final answer catches hallucination.Adaptive RAG, so each question type gets its own retrieval strategy.Corrective RAG-style grading that sends cases below the threshold to a human queue with the retrieved context attached.Plain RAG with a larger top-k, since the corpus will eventually document the new codes.
6. 6. On a shared platform, a restricted user keeps getting thin or empty context even though their own documents contain good answers. The pipeline retrieves and reranks a top 6 across the whole corpus, then removes the documents the user cannot see. What should change?

   Enforce access first: build the keyword index over the user's authorised pool and rerank after access enforcement.Raise the reranker's output from 6 to 20 so that more permitted documents survive the post-filter.Swap the LLM reranker for a local cross-encoder so it can score more candidates.Add multi-query rewriting so retrieval finds more paraphrases of the question.
7. 7. A team runs six retrieval strategies against a 22-document golden set. Every strategy scores recall@k of 1.00, and HyDE has a slightly higher mean reciprocal rank than the rest. What is the right conclusion?

   Ship HyDE, since it has the highest mean reciprocal rank of the six.Ship the combination of every strategy, since combining them cannot lower recall.Drop the golden set, since a metric that every strategy maxes out is useless.Treat the differences as noise, keep the harness and rerun it on a realistically sized corpus before promoting any strategy.
8. 8. You are adding BM25 alongside dense search so that a ticket ID such as MRD-5031 stops returning the MRD-5030 guide. The dense side already searches with several rewrites of the question. Which design keeps exact IDs findable?

   Add the BM25 score directly to the dense similarity score so an exact match lifts the total.Keep hyphenated IDs as single tokens in the keyword index, and send BM25 only the original question and any sub-questions.Split IDs on the hyphen so that MRD and 5031 each match, and feed every rewrite to BM25 as well.Skip BM25 and move to a stronger embedding model that separates similar codes.