What is Qdrant's Discovery/Recommendation API used for?

Beyond a standard single-query similarity search, Qdrant provides a Recommendation API and a Discovery API for scenarios where relevance is better expressed through multiple example points — some to move toward, some to move away from — rather than a single query vector.

results = client.recommend(
    collection_name="products",
    positive=[101, 205],  # point IDs the user liked
    negative=[42],        # a point ID the user disliked
    limit=10,
)

The Recommendation API takes lists of positive and negative example point IDs (or raw vectors) and finds points similar to the positive examples while being dissimilar to the negative ones, which is a natural fit for "more like this, but not like that" scenarios common in recommendation systems — suggesting products similar to ones a user has purchased, while avoiding ones similar to items they've explicitly rejected.

The Discovery API goes a step further, letting a search be guided by a broader set of context pairs (each pair expressing "prefer results closer to A than to B") alongside an optional primary target, which supports more nuanced exploration than a single positive/negative example, such as guiding a search toward a general vibe or theme defined by multiple contrasting reference points rather than one single anchor.