/match returns results below threshold unless deprecated cutoff is explicitly passed

[santialbo]

When calling /match with a threshold higher than the default, the response still includes results scored below threshold unless the deprecated cutoff parameter is explicitly provided and set to the same value.

threshold: float = Query(
    default=settings.SCORE_THRESHOLD,
    title="Score threshold for results to be considered matches",
),
cutoff: float = Query(
    deprecated=True,
    default=settings.SCORE_THRESHOLD,
    title="Deprecated, use `threshold` instead. Lower bound of score for results to be returned at all",
),
...
cutoff = min(cutoff, threshold)
...
total, scored = await score_results(
    algorithm_type,
    entity,
    ents,
    threshold=threshold,
    cutoff=cutoff,
    limit=limit,
    config=ScoringConfig(weights=match.weights, config=match.config),
)

As a result, a higher threshold does not tighten the lower-bound filter unless cutoff is also raised. Given that cutoff is marked deprecated, this is surprising behavior.

Expected behavior
Setting a higher threshold should be sufficient to prevent results below that threshold from appearing in the response, even if cutoff is not passed.

[leonhandreke]

Hey Santi,

so the idea is that

• score < cutoff -> not returned
• cutoff <= score < threshold -> returned with match = false
• threshold < score -> returned with match = true

What do you think should happen? Should the cutoff value follow threshold unless cutoff is explicitly set?

[santialbo]

I guess the deprecation message for cutoff confused me. If cutoff is being deprecated how will it work on upcoming versions? Should we filter the results ourselves?

[leonhandreke]

Sorry for the late reply, missed this one... Yeah, it's a bit of a difficult situation we've gotten ourselves into, because the parameters are named in a fairly confusing way. The only thing that is controlled by threshold is the match: true/false. Our idea with deprecation was that this was never very useful and that your logic will likely be more complex anyway, like:

if score > 0.95:
  sound_the_alarm()
elif score > 0.9:
  button_color = red
elif score > 0.8:
  button_color = orange
elif score > 0.7:
  button_color = yellow

Hope that makes sense!