Recommendation System Design Interview Questions (2026): 14 Questions

A 5-step way to answer any recommender design question

Clarify the goal. What to optimize (engagement, purchases, long-term satisfaction) and the scale.
Define metrics. Offline ranking metrics plus online business and guardrail metrics.
Design the funnel. Candidate generation, ranking and re-ranking with rough candidate counts.
Choose data and features. User, item and context features plus labels from interactions.
Plan serving and feedback. Latency budget, freshness, logging and bias correction.

Contents: Architecture (questions 1 to 4) · Data and metrics (questions 5 to 8) · Cold start and serving (questions 9 to 14)

Architecture (questions 1 to 4)

1. Why do recommenders use multiple stages?

Scoring every item with a heavy model is too slow for large catalogs.
Candidate generation cheaply narrows millions of items to hundreds or thousands.
Ranking scores candidates with richer features.
Re-ranking applies diversity, freshness and business rules.

Common pitfall: Proposing a single model over the entire catalog.

2. What is a two-tower model?

One tower embeds the user and context; the other embeds the item.
Relevance is the dot product or cosine of the two embeddings.
Item embeddings are precomputed and searched with approximate nearest neighbor indexes.
It is trained on interactions, often with in-batch negatives.

Common pitfall: Forgetting that it cannot use rich user-item cross features.

3. What does the ranking model do?

It predicts the probability of target actions for each candidate, such as click, like or purchase.
It uses user, item, context and cross features.
Multi-task models predict several actions and combine them with weights.
It runs on hundreds of candidates within the latency budget.

Common pitfall: Optimizing only clicks, which can reward clickbait.

4. What happens in re-ranking?

Enforce diversity so results are not all the same type.
Apply freshness, fairness and business rules, such as removing already purchased items.
Deduplicate and enforce policy filters.
Measure the effect on engagement and satisfaction.

Common pitfall: Skipping re-ranking and serving a monotonous list.

Data and metrics (questions 5 to 8)

5. Which metrics do you use?

Offline: recall@k for candidate generation; nDCG, AUC or log loss for ranking.
Online: engagement, conversion, retention and session length.
Guardrails: diversity, complaint or hide rates and latency.
Confirm with A/B tests.

Common pitfall: Using only offline metrics.

6. What is position bias and how do you handle it?

Items shown higher get more clicks regardless of relevance.
Training on raw clicks teaches the model to repeat past rankings.
Mitigate with position as a training feature set to a constant at serving, inverse propensity weighting or randomized exploration data.
Evaluate with unbiased or randomized traffic when possible.

Common pitfall: Treating every non-click as a negative label.

7. Explicit versus implicit feedback: which do you use?

Explicit feedback (ratings) is clear but sparse.
Implicit feedback (clicks, watch time, purchases) is plentiful but noisy.
Most production systems rely on implicit signals with careful label definitions.
Define negatives carefully, for example items shown but skipped.

Common pitfall: Using watch time without accounting for content length.

8. How do you avoid feedback loops?

The system trains on data it influenced, which can narrow what users see.
Add exploration to show new or diverse items.
Monitor catalog coverage and popularity concentration.
Balance short-term clicks with long-term satisfaction metrics.

Common pitfall: Ignoring popularity bias until the catalog tail disappears.

Cold start and serving (questions 9 to 14)

9. How do you handle new users?

Use onboarding questions, popular items and context such as location and device.
Update recommendations quickly from early session behavior.
Use exploration to learn preferences.
Measure new-user retention separately.

Common pitfall: Showing nothing personalized until history accumulates.

10. How do you handle new items?

Use content features (text, images, metadata) so items get embeddings without interactions.
Reserve exploration traffic for new items.
Boost freshness for time-sensitive content.
Track how fast new items get exposure.

Common pitfall: Relying only on collaborative filtering.

11. How do you serve recommendations within latency limits?

Precompute item embeddings and maintain an ANN index.
Cache user features and popular results.
Limit ranking to a bounded candidate set.
Set a per-stage latency budget and fallbacks.

Common pitfall: Fetching many features synchronously from slow stores.

12. How do you keep recommendations fresh?

Stream recent interactions into user features.
Retrain models regularly and update embeddings.
Use real-time signals such as session history in ranking.
Monitor model and feature staleness.

Common pitfall: Retraining monthly for a fast-moving content feed.

13. Where do LLMs fit in recommendation systems today?

Generating item descriptions, tags or embeddings for content understanding.
Understanding natural language queries or user intent.
Explaining recommendations.
Classic retrieval and ranking models still handle most large-scale scoring because of cost and latency.

Common pitfall: Proposing to call an LLM for every candidate item.

14. Design a "people you may know" feature. What is different?

Candidates come from the social graph, such as friends of friends and shared groups.
Features include mutual connections and interaction history.
Privacy and safety rules limit what can be suggested.
Metrics include accepted requests and complaint rates.

Common pitfall: Ignoring privacy constraints.

Go deeper on related questions

FAQ

Which recommender prompts are most common?

Feed ranking, product recommendations, video recommendations and "people you may know".

Do I need deep learning knowledge?

Know two-tower retrieval and a deep ranking model, but explain why gradient-boosted trees can still be a strong ranking baseline.

How detailed should my numbers be?

Rough orders of magnitude are enough, such as millions of items narrowed to hundreds of candidates.

How do I practice?

Design three recommenders out loud in 40 minutes each, using the same five-step structure.

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