AI Engineer Interview Questions / Interview topics

Hybrid Search and Reranking Interview Questions (2026): 14 Questions

Hybrid search and reranking questions test whether you can improve retrieval beyond a single vector search. Interviewers ask when keyword and dense retrieval each fail, how to fuse their results, what a reranker adds, how many candidates to rerank and how to measure the gain. Use the framework below and connect every choice to recall, precision at the top and latency.

Get the playbook Free 50 questions

A 5-step way to answer any hybrid search question

  1. Name the failure. Exact terms missed by dense search, or meaning missed by keyword search.
  2. Combine retrievers. BM25 or learned sparse plus dense embeddings.
  3. Fuse results. Reciprocal rank fusion or tuned weighted scores.
  4. Rerank. Cross-encoder over a bounded candidate set.
  5. Measure. Recall@k, nDCG and latency on a labeled query set.

Contents: Retrieval methods (questions 1 to 4) · Fusion (questions 5 to 7) · Reranking (questions 8 to 11) · Evaluation and operations (questions 12 to 14)

Retrieval methods (questions 1 to 4)

1. How does BM25 work?

  • It scores documents by term frequency, inverse document frequency and document length normalization.
  • Term frequency saturates, so repeating a word has diminishing effect.
  • It excels at exact matches such as names, codes and rare terms.
  • It does not understand synonyms or paraphrases.

Common pitfall: Calling BM25 outdated when it is a strong baseline.

2. Where does dense retrieval fail?

  • Exact identifiers, error codes, product SKUs and rare terms.
  • Domains far from the embedding model's training data.
  • Very short or ambiguous queries.
  • Hybrid search covers many of these cases.

Common pitfall: Assuming embeddings capture every exact match.

3. What is hybrid search?

  • Running keyword (sparse) and semantic (dense) retrieval and combining the results.
  • It improves recall across query types.
  • Many vector databases and search engines support it natively.
  • Requires a fusion method and evaluation.

Common pitfall: Using hybrid search without measuring whether it helps.

4. What is learned sparse retrieval such as SPLADE?

  • A model produces sparse term weights, including expansion terms not in the text.
  • It keeps the efficiency of inverted indexes with some semantic matching.
  • It can replace or complement BM25 in hybrid setups.
  • It needs model inference at indexing and query time.

Common pitfall: Confusing it with dense embeddings.

Fusion (questions 5 to 7)

5. What is reciprocal rank fusion?

  • Each document's score is the sum over lists of 1 / (k + rank).
  • It uses ranks, so it ignores incompatible score scales.
  • The original paper used k = 60.
  • It is a robust default that needs no training.

Common pitfall: Adding raw BM25 and cosine scores directly.

6. When would you use weighted score fusion instead?

  • When you have labeled data to tune weights.
  • Normalize scores first, for example min-max per query.
  • Tuned weights can beat RRF on your domain.
  • Re-tune when retrievers or data change.

Common pitfall: Using fixed weights picked without data.

7. How many candidates should each retriever return?

  • Enough for good recall before reranking, often tens to a few hundred in total.
  • More candidates improve recall but raise reranking cost.
  • Tune using recall@k curves on labeled queries.
  • Deduplicate overlapping results.

Common pitfall: Retrieving only the final k from each retriever.

Free

Get the free 50 RAG and LLM interview questions

Join the voluntary email list and open the free set of 50 RAG and LLM interview questions with short model answers. You will also get occasional new question sets. No spam, and you can unsubscribe at any time by replying "unsubscribe" to any email.

We only use your email for this list. No results are promised; this is free practice material.

Reranking (questions 8 to 11)

8. What is a cross-encoder reranker?

  • It reads the query and a document together and outputs a relevance score.
  • It captures interactions that separate embeddings miss.
  • It is much slower than bi-encoder similarity, so it runs on a short list.
  • It usually improves precision at the top results.

Common pitfall: Running a cross-encoder over the whole corpus.

9. Bi-encoder versus cross-encoder: what is the trade-off?

  • Bi-encoders embed query and documents separately; documents can be precomputed.
  • Cross-encoders score pairs jointly and are more accurate per pair.
  • Use bi-encoders for retrieval and cross-encoders for reranking.
  • Latency scales with the number of reranked pairs.

Common pitfall: Saying cross-encoders can be used for first-stage retrieval at scale.

10. Can an LLM be used as a reranker?

  • Yes, by scoring or ordering candidates with a prompt.
  • It can be accurate but is slower and more expensive.
  • Use it for small candidate sets or offline labeling.
  • Check position bias in listwise prompts.

Common pitfall: Ignoring cost when proposing LLM reranking for every query.

11. How do you fit reranking into a latency budget?

  • Limit the candidate count and truncate long documents.
  • Use a smaller reranker model or GPU batching.
  • Cache results for repeated queries.
  • Measure p95 latency end to end.

Common pitfall: Reranking too many long documents.

Evaluation and operations (questions 12 to 14)

12. How do you evaluate a hybrid search change?

  • Build a labeled query set with relevant documents.
  • Compare recall@k, MRR or nDCG before and after.
  • Break results down by query type, such as exact-match versus conceptual.
  • Confirm with online metrics.

Common pitfall: Evaluating with a few hand-picked queries.

13. How do you debug a query that returns bad results?

  • Check each retriever's results separately.
  • Check fusion and reranker scores for the expected document.
  • Inspect chunking and metadata filters.
  • Add the query to the test set.

Common pitfall: Changing several components at once.

14. What operational issues come with hybrid search?

  • Two indexes to keep in sync with updates and deletions.
  • Tokenization and language settings for the keyword index.
  • Embedding model versions for the dense index.
  • Monitoring of latency and result quality.

Common pitfall: Updating the vector index but forgetting the keyword index.

Go deeper on related questions

FAQ

Is hybrid search always better?

Often, but not always. Measure on your queries; for purely conceptual queries dense search alone may be enough.

What is a typical rerank setup?

Retrieve a few dozen to a few hundred candidates with hybrid search, rerank with a cross-encoder and keep the top few for the prompt.

Do I need to know BM25 parameters?

Knowing that k1 controls term frequency saturation and b controls length normalization is enough for most interviews.

How do I practice?

Index a small corpus with BM25 and embeddings, label 30 queries and compare recall@k for each setup.

The PlaybookWant the full method?

The 0 to 1 AI Engineer Interview Playbook covers answer frameworks, system-design walkthroughs and an 8-week plan. $19.99 on Gumroad.

Get the playbook

All interview topics · Prep guides · Question deep dives · Free 50 RAG and LLM questions