Automatic suggestion detection threshold

[tomkolp]

Has the automatic suggestion threshold been changed recently? Now most automatic suggestions are only 100%, rarely less. We used these smaller thresholds a lot. The longer the text, the worse it is.

[nijel]

Yes, the PostgreSQL similarity threshold calculation has been adjusted some time ago to do only closer matches:

weblate/weblate/memory/models.py

Lines 62 to 81 in 68df5a1

	def threshold_to_similarity(self, text: str, threshold: int) -> float:
	"""
	Convert machinery threshold into PostgreSQL similarity threshold.
	Machinery threshold typical values:
	- 75 machinery
	- 80 automatic translation (default value)
	- 10 search
	PostgreSQL similarity threshold needs to be higher to avoid too slow
	queries.
	"""
	length = len(text)
	base = 0.172489 * math.log(threshold) + 0.207051
	bonus = 7.03436 * math.exp(-6.07957 * base)
	length_bonus = (0.0733418 * math.log(length) + 0.324497) * bonus
	return max(0.6, min(1.0, round(base + length_bonus, 3)))

If using automatic translation, you can configure the similarity threshold in it.

[tomkolp]

Currently it is 75 and should I change this value to a smaller one? How long was it before?

Do I need to make changes inside the weblate-docker-weblate-1 container? Do you have to do it again after each update?

Is it possible to enter these settings into the gui? It would be great if you could change this value for a specific component.

[nijel]

There is no way to configure what "Automatic suggestions" tab does. But when using add-on or invoke automatic translation in the UI, you define your own threshold there.