Weighted vectors for recommendation api

[fungilation]

Looking at https://qdrant.tech/documentation/concepts/search/?selector=aHRtbCA%2BIGJvZHkgPiBkaXY6bnRoLW9mLXR5cGUoMSkgPiBzZWN0aW9uID4gZGl2ID4gZGl2ID4gZGl2ID4gYXJ0aWNsZSA%2BIGgyOm50aC1vZi10eXBlKDgp#recommendation-api

First, a clarification question in the example:

  "negative": [718],
  "positive": [100, 231]

Are the the example numbers in the arrays the id of "points" in a Qdrant collection, and not vectors themselves?

In building recommendation systems, I'm strongly leaning towards using Qdrant with the simplicity of this api. However, a little too simple: I need a way to weigh different input (positive) vectors in how the results list are scored.

I could the weighting by issuing multiple qdrant recommendation api calls, aggregate the multiple results, and recombine them with weights. But it'd be a great improvement, for both DX and minimizing network round trips, to have weighted vector ids. Example:

  "negative": [718],
  "positive": [{100: 1.0}, {231: 0.3}]

where 1.0, 0.3 are the weights, and they are floats in range of 0-1.

Thoughts on usefulness and complexity of implementing this welcome!