41 Search and Recommendation System Interview Questions

All 41 questions

What are embeddings?
Asked in 4 of 320 interviews · last asked Oct 19 · Concept · Junior
Show answer
Embeddings are vectors of numbers that represent items such as words, images, or users. They are usually learned from data, so items with similar meaning or behavior have nearby vectors. They are used in semantic search, recommendation systems, and as inputs to machine learning models.
- Embeddings capture useful relationships, not just unique labels for items.
- Similarity can be measured with cosine similarity, dot product, or distance, depending on the model.
- What counts as similar depends on the training data and task.
Code
```
A search model embeds a question and a set of documents.
It compares their vectors to find relevant documents.
This can find matches even when the exact words differ.
```
What the interviewer checks: The interviewer is checking whether you understand what embeddings represent and how they support similarity-based tasks.
Other ways it was asked
- “What is your understanding of embeddings?”
- “What are embeddings, and when are they used?”
- “What are embeddings, and where are they used?”
When would you use supervised rather than unsupervised learning for a recommendation model?
Asked once · last asked Sep 28 · Concept · Mid-level
Other ways it was asked
- “How would you use supervised learning to train a recommendation model?”
- “Why might supervised learning be a better choice than unsupervised learning for a recommendation model?”
- “How do supervised learning and reinforcement learning differ when building a recommendation model?”
Which library would you use to calculate cosine similarity and match items?
Asked once · last asked Aug 19 · Tooling · Mid-level
How would you evaluate a hybrid BM25 and k-nearest-neighbor vector-search approach, and which metrics would you use?
Asked once · last asked Sep 7 · Concept · Senior
Other ways it was asked
- “How would you test whether adding k-nearest-neighbor vector search to BM25 improves performance compared with a BM25 baseline?”
- “How would you evaluate a hybrid BM25 and vector-search approach, and how would you choose the weights for each retrieval method?”
How do embeddings work in a computer system, and why are they used?
Asked once · last asked Apr 16 · Concept · Junior
Comparing the last 90 days of recommendations with customer behavior, how would you define success and a reported 30% failure rate precisely enough to identify patterns and verify that rule changes fixed the problem rather than shifted it elsewhere?
Asked once · last asked May 7 · Scenario · Senior
Other ways it was asked
- “When you compare the last 90 days of recommendations with customer behavior, what exactly are you measuring as success? How do you define the 30% failure rate precisely enough to identify a pattern, and how do you know your rule changes fixed it rather than moving the problem somewhere else?”
What data would you use to build an offer-to-user ranking model, and how would you distinguish available features from assumptions?
Asked once · last asked May 13 · Scenario · Senior
Which retrieval metrics can help assess whether retrieved chunks are complete and relevant?
Asked once · last asked May 13 · Concept · Mid-level
How would you evaluate a model's accuracy and retrieval quality using metrics such as precision, recall, or mean average precision?
Asked once · last asked Jun 12 · Scenario · Senior
How would you design an offline evaluation that reliably predicts the online impact of a ranking or recommendation model, including how you would avoid leakage and handle delayed feedback?
Asked once · last asked Jun 15 · Scenario · Senior
How do you compute and apply propensity scores for the behavior policy in inverse propensity scoring, and how do you control variance from very small propensities?
Asked once · last asked Jun 15 · Concept · Senior
How would you estimate the true relevance of items from biased click logs, for example, by accounting for position bias in an offline evaluation?
Asked once · last asked Jun 15 · Scenario · Senior
How would you handle variance and exploding weights in inverse propensity scoring, for example, when propensities are very small at low ranks, while keeping the estimator reliable?
Asked once · last asked Jun 15 · Concept · Senior
When you suspect propensity depends on context such as device, time, or user segment, how would you design the logging policy and estimator so the positivity assumption holds for all candidate items?
Asked once · last asked Jun 15 · Scenario · Senior
What is Precision@k, and how is it used to evaluate a RAG system?
Asked once · last asked Jun 17 · Concept · Mid-level
How would you investigate an article recommendation system to validate its shortcomings and identify where changes are most needed?
Asked once · last asked Jun 17 · Scenario · Mid-level
Which data elements, metadata, or relationship features would you add or enhance to strengthen knowledge-aware recommendations in new regions?
Asked once · last asked Jun 17 · Scenario · Senior
How would you evaluate recommendations to ensure they meet performance expectations in different regions?
Asked once · last asked Jun 17 · Scenario · Mid-level
Which metrics do you use to evaluate retrieval performance, especially when the classes are imbalanced?
Asked once · last asked Jul 16 · Concept · Senior
Besides recall and precision, what classical retrieval metrics would you use to evaluate retrieved chunks?
Asked once · last asked Jul 20 · Concept · Mid-level
Other ways it was asked
- “Besides recall and precision, what classical retrieval metrics did you use to evaluate retrieved chunks?”
In retrieval evaluation with top-10 results, how do you calculate precision, and what counts as a true positive?
Asked once · last asked Jul 20 · Concept · Mid-level
How would you build a recommendation system that handles sparse viewing history and new users?
Asked once · last asked Aug 1 · Design · Mid-level
Other ways it was asked
- “Suppose you are building a recommendation system for Amazon Prime Video using user viewing history, search queries, and demographic data, but a significant portion of viewing history is missing or sparse, especially for newer users. How would you design a machine learning pipeline to handle this data sparsity issue?”
- “Which algorithms, data analysis steps, or engineering techniques would you evaluate to address data sparsity in a recommendation system?”
What metrics would you use to diagnose and improve a semantic video-search model and its architecture?
Asked once · last asked Sep 7 · Concept · Senior
How would you calculate precision and query accuracy for a search system?
Asked once · last asked Sep 7 · Concept · Mid-level
How would you judge search success when users click lower-ranked results rather than the first result?
Asked once · last asked Sep 7 · Scenario · Senior
Other ways it was asked
- “How would you determine whether a search was successful when a user clicks the third or fourth result rather than the first?”
- “Should search-result relevance be defined solely by whether the user clicks the first result?”
How would you evaluate the effectiveness of ranking videos using their content context rather than metadata alone?
Asked once · last asked Sep 7 · Concept · Senior
How does cosine similarity work on two numeric vectors?
Asked once · last asked Sep 15 · Concept · Junior
How would you design an end-to-end semantic search system, from user query to ranked results, and measure its improvement?
Asked once · last asked Sep 16 · Design · Senior
How would you build a cross-sell, upsell, and next-best-recommendation model for a B2B business?
Asked once · last asked Sep 22 · Scenario · Senior
Would you use large language models, small language models, or classical machine learning to build a B2B cross-sell and upsell recommendation model?
Asked once · last asked Sep 22 · Scenario · Senior
What are recall and mean reciprocal rank (MRR), and how are they used to evaluate retrieval?
Asked once · last asked Oct 7 · Concept · Mid-level
In duplicate detection, what does a cosine similarity score of 0.91 indicate?
Asked once · last asked Oct 20 · Scenario · Mid-level
When setting up search terms or TAR for first-pass responsiveness, how do you validate recall and avoid over-excluding relevant documents before privilege review?
Asked once · last asked Oct 21 · Scenario · Senior
How would you measure whether a retriever finds the correct passages?
Asked once · last asked Oct 22 · Concept · Mid-level
What are embeddings, and why are they used in AI applications?
Asked once · last asked Oct 23 · Concept · Junior
What algorithms can be used in an AI-powered product search system?
Asked once · last asked Mar 27 · Concept · Mid-level
Other ways it was asked
- “How can AI be used to improve a product search system?”
- “What techniques or modules can be used to implement AI-powered product search?”
What does precision measure when evaluating a RAG pipeline?
Asked once · last asked Sep 10 · Concept · Junior
Other ways it was asked
- “How do you calculate precision when evaluating a RAG pipeline?”
What is a false positive in RAG evaluation?
Asked once · last asked Sep 10 · Concept · Junior
Other ways it was asked
- “What is a false negative in RAG evaluation?”
What is an embedding?
Asked once · last asked Aug 7 · Concept · Junior
Other ways it was asked
- “What is embedding?”
What does the embedding model do?
Asked once · last asked Sep 15 · Concept · Junior
How does AI help real-estate leads find properties?
Asked once · last asked Oct 23 · Concept · Mid-level
Other ways it was asked
- “From a technical perspective, how can AI help real-estate leads find suitable properties?”

How this list is made

Every question here was asked by a real interviewer in a real job interview that a candidate ran MeetAssist in. After the call, the interviewer's questions are pulled from the transcript. Practice sessions, tests, exams and anything that was not a job interview are dropped.

Questions are rewritten so they make sense on their own: names of people, employers and clients are removed, and speech-to-text mistakes are fixed. Small talk, logistics and questions about one candidate's own résumé are left out. Rephrasings of the same question are merged, and the count is the number of different interviews it came up in.

The interviews come from candidates who use MeetAssist, so they lean towards software, data and AI roles: the shares describe these interviews, not every Machine Learning interview. Answers to the most asked questions are written and checked by an editor; a "common mistake" is shown only when reviews of real answers in these interviews show it, with the count.

320 interviews, 221 candidates, Nov 2025 to Oct 2026. Updated every month with new interviews.

All 41 questions

What are embeddings?

When would you use supervised rather than unsupervised learning for a recommendation model?

Which library would you use to calculate cosine similarity and match items?

How would you evaluate a hybrid BM25 and k-nearest-neighbor vector-search approach, and which metrics would you use?

How do embeddings work in a computer system, and why are they used?

Comparing the last 90 days of recommendations with customer behavior, how would you define success and a reported 30% failure rate precisely enough to identify patterns and verify that rule changes fixed the problem rather than shifted it elsewhere?

What data would you use to build an offer-to-user ranking model, and how would you distinguish available features from assumptions?

Which retrieval metrics can help assess whether retrieved chunks are complete and relevant?

How would you evaluate a model's accuracy and retrieval quality using metrics such as precision, recall, or mean average precision?

How would you design an offline evaluation that reliably predicts the online impact of a ranking or recommendation model, including how you would avoid leakage and handle delayed feedback?

How do you compute and apply propensity scores for the behavior policy in inverse propensity scoring, and how do you control variance from very small propensities?

How would you estimate the true relevance of items from biased click logs, for example, by accounting for position bias in an offline evaluation?

How would you handle variance and exploding weights in inverse propensity scoring, for example, when propensities are very small at low ranks, while keeping the estimator reliable?

When you suspect propensity depends on context such as device, time, or user segment, how would you design the logging policy and estimator so the positivity assumption holds for all candidate items?

What is Precision@k, and how is it used to evaluate a RAG system?

How would you investigate an article recommendation system to validate its shortcomings and identify where changes are most needed?

Which data elements, metadata, or relationship features would you add or enhance to strengthen knowledge-aware recommendations in new regions?

How would you evaluate recommendations to ensure they meet performance expectations in different regions?

Which metrics do you use to evaluate retrieval performance, especially when the classes are imbalanced?

Besides recall and precision, what classical retrieval metrics would you use to evaluate retrieved chunks?

In retrieval evaluation with top-10 results, how do you calculate precision, and what counts as a true positive?

How would you build a recommendation system that handles sparse viewing history and new users?

What metrics would you use to diagnose and improve a semantic video-search model and its architecture?

How would you calculate precision and query accuracy for a search system?

How would you judge search success when users click lower-ranked results rather than the first result?

How would you evaluate the effectiveness of ranking videos using their content context rather than metadata alone?

How does cosine similarity work on two numeric vectors?

How would you design an end-to-end semantic search system, from user query to ranked results, and measure its improvement?

How would you build a cross-sell, upsell, and next-best-recommendation model for a B2B business?

Would you use large language models, small language models, or classical machine learning to build a B2B cross-sell and upsell recommendation model?

What are recall and mean reciprocal rank (MRR), and how are they used to evaluate retrieval?

In duplicate detection, what does a cosine similarity score of 0.91 indicate?

When setting up search terms or TAR for first-pass responsiveness, how do you validate recall and avoid over-excluding relevant documents before privilege review?

How would you measure whether a retriever finds the correct passages?

What are embeddings, and why are they used in AI applications?

What algorithms can be used in an AI-powered product search system?

What does precision measure when evaluating a RAG pipeline?

What is a false positive in RAG evaluation?

What is an embedding?

What does the embedding model do?

How does AI help real-estate leads find properties?

More Machine Learning interview questions

How this list is made

Walk into your next interview ready.