Lower automatic suggestions threshold in case of no results

[strepon]

Describe the problem

In automatic suggestions, the threshold to show them (75 as written here) is sometimes too high and useful suggestions are omitted. As a workaround, user can enter the string into the "Translation memory" field and see the search results where the threshold is lower (10).

Describe the solution you would like

For me, it would make sense to lower the threshold automatically until a first result is found.

As an alternative, a more convenient and helpful UI could be introduced - e.g. a button which would perform the search without the need to copypaste the string manually (or a button which would copy the string to the "Translation memory" field).

Describe alternatives you have considered

No response

Screenshots

An example from The Document Foundation instance of Weblate:

There are strings:

• original: "This setting enables you to export the document as a .pdf file containing two file formats: PDF and ODF."
• and newer "This setting enables you to export the document as a .pdf file containing two file formats: PDF and ODF as an attachment."

differing only in the last "as an attachment" text.

However, the original string is not shown in the suggestions for the newer one (but translators would expect it):

If using the search, the string is there:

(Which corresponds with the thresholds: for these strings, thresholds 75 and 10 mean similarities 97 % and 62 %; the suggestion has 85 %).

Additional context

No response

[tomkolp]

We have this problem too, especially with very long strings. A manual search button with a lower threshold would be ideal.

[nijel]

#13841 should make the matching better for longer strings, even though it was more targeted at strings containing markup. I see TDF is still on 5.9, so the improvement might be there once the server is upgraded to 5.10.

@tomkolp did you observe any change in this recently? You probably already run this.

[strepon]

OK, let's wait to see this in 5.10 (I didn't realize that #13841 affects all strings with sentences, it would be great if it helps).

[nijel]

It does as it removes all punctuation and spaces from the calculation. This will always be a trade-off between results quality and performance. In the long term, I think we will have to migrate translation memory outside the database and use OpenSearch for it (or something entirely different, I haven't really investigated other options).

[strepon]

After testing it in 5.10, I am confused - it seems the suggestions are not provided even for high similarities, e.g. in this case - I think this worked previously: