What is Qdrant's recommend query mode, and how is it different from plain similarity search?
Recommendation queries
A recommend query is useful when the desired item is better described by examples than by a single explicit query vector. You provide positive examples and optionally negative examples, and Qdrant derives a recommendation direction from those points.
That differs from ordinary similarity search, where the query is a vector supplied by the caller and Qdrant ranks points by distance to that vector. Recommendation queries let the retrieval intent be expressed as 'find items like these, but unlike those' without the application necessarily constructing the embedding itself.
The trade-off is expressiveness versus predictability. A recommendation can be useful for personalization and 'more like this' behavior, while direct vector search is simpler when the application already has a well-defined query embedding.
A common misconception is that recommend is a completely separate similarity algorithm. It still relies on vector-space similarity; the distinction is how the query representation is constructed from positive and negative examples. Supported recommendation semantics can evolve with Qdrant releases.
- 1
Recommendation queries use positive and optional negative examples
- 2
Plain similarity search uses an explicit query vector
- 3
Recommend is useful for 'more like these' and preference-driven retrieval
- 4
Recommendation still operates in the configured vector space
Scenario Questions
0-2 years experience
- 1A user likes three products and dislikes one. Which Qdrant query pattern could express 'find products like the liked ones but unlike the disliked one'?
- 2Why might recommend be more convenient than manually averaging several product embeddings?
2-5 years experience
- 1A recommendation system returns items that are similar to positive examples but still resemble a disliked category. How would you investigate the positive and negative example selection?
- 2A product team wants 'more like this item' and already has the item's vector. Would you use recommend or ordinary similarity search? Explain the choice.
5-8 years experience
- 1You need personalized recommendations based on several positive interactions and a small set of explicit dislikes. How would you design the query and evaluate whether the negative examples improve ranking?
- 2Recommendation quality varies significantly by user because some users have only one positive example. How would you handle sparse preference signals?
8+ years experience
- 1You are designing a recommendation platform where preference signals vary from explicit likes to implicit behavior. How would you decide when to use Qdrant recommend versus an external ranking model?
- 2A business stakeholder wants negative examples to guarantee certain categories never appear. Would you use recommend for that requirement? What would you use instead for a hard constraint?
Follow-up Questions
- How do negative examples affect a recommendation query?
- When is a normal vector query preferable to recommend?