Local mode: recommend/discovery/context/feedback queries mis-order and mis-threshold on Euclid/Manhattan

[winklemad]

Current Behavior

Recommend / discovery / context / relevance-feedback queries score points from the internal core distance, which is oriented so a higher score is always better regardless of the collection metric. On Euclid/Manhattan collections in local mode, two paths in LocalCollection.search treat those scores using the raw distance order instead, producing wrong results.

1. Relevance-feedback returns the farthest points first (reversed order).

c = QdrantClient(":memory:")
c.create_collection("rf", vectors_config=models.VectorParams(size=2, distance=models.Distance.EUCLID))
c.upsert("rf", points=[
    models.PointStruct(id=1, vector=[0.1, 0.0]),   # nearest to target [0, 0]
    models.PointStruct(id=5, vector=[5.0, 0.0]),
    models.PointStruct(id=9, vector=[9.0, 0.0]),   # farthest
])
c.query_points("rf", query=models.RelevanceFeedbackQuery(relevance_feedback=models.RelevanceFeedbackInput(
    target=[0.0, 0.0],
    feedback=[models.FeedbackItem(example=[3.0, 0.0], score=0.5),
              models.FeedbackItem(example=[4.0, 0.0], score=0.5)],
    strategy=models.NaiveFeedbackStrategy(naive=models.NaiveFeedbackStrategyParams(a=1.0, b=1.0, c=1.0)),
)), limit=3).points
# ids -> [9, 5, 1]   (expected [1, 5, 9])

2. score_threshold drops all results for the whole higher-is-better family (recommend/discovery/context/feedback).

c.create_collection("t", vectors_config=models.VectorParams(size=2, distance=models.Distance.EUCLID))
c.upsert("t", points=[models.PointStruct(id=i, vector=[float(i), 0.0]) for i in (1, 2, 3, 4)])
c.query_points("t",
    query=models.RecommendQuery(recommend=models.RecommendInput(
        positive=[[1.0, 0.0]], negative=[], strategy=models.RecommendStrategy.BEST_SCORE)),
    limit=10, score_threshold=0.13).points
# returns []   (expected the points scoring >= 0.13)

Expected Behavior

Feedback results ordered nearest-first ([1, 5, 9]), and score_threshold keeps the points scoring at/above the threshold for the recommend/discovery/context/feedback family.

Root cause

Both the result ordering and the score_threshold check derive their direction from distance_to_order(distance) (the raw distance order). For the recommend/discovery/context/feedback family the scores are higher-is-better on every metric, so on Euclid/Manhattan the ordering special-case must apply (it omitted NaiveFeedbackQuery) and the threshold direction must match — instead the threshold followed the raw order and inverted.

I have a fix that derives a single higher-is-better flag from the query type and uses it for both the ordering and the threshold, plus regression tests over both metrics.