Re: Sorting by respecting diacritics/accents

On Fri, 2025-07-25 at 13:05 +0300, JānisE wrote:
> I seem to not be able to get PostgreSQL to sort rows by a string column respecting the diacritics.
>
> I read [1] that it's possible to define a custom collation having collation strength "ks"
> set to "level2", which would mean that it's accent-sensitive.
>
> However, when I try to actually sort using that collation, the order seem to be accent-insensitive.
>
> For example:
>
>  CREATE TABLE test (string text);
>  INSERT INTO test VALUES ('bar'), ('bat'), ('bär');
>  CREATE COLLATION "und1" (provider = icu, deterministic = false, locale = 'und-u-ks-level1');
>  CREATE COLLATION "und2" (provider = icu, deterministic = false, locale = 'und-u-ks-level2');
>  CREATE COLLATION "und3" (provider = icu, deterministic = false, locale = 'und-u-ks-level3');
>  SELECT * FROM test ORDER BY string collate "und1";
>  SELECT * FROM test ORDER BY string collate "und2";
>  SELECT * FROM test ORDER BY string collate "und3";
>
> All three collations give me the same order: bar < bär < bat, although an accent-sensitive
> order would be bar < bat < bär
>
> This does lose "bär", meaning that those strength levels do have some kind of an effect on "DISTINCT":
> SELECT DISTINCT string COLLATE "und1" FROM test;
>
> But it's not working on "ORDER BY".
>
> Do I misunderstand the collation capabilities? Is there a way to actually get an accent-sensitive order?

Yes, I thing you misunderstand what "accent sensitive" means.
It means that 'bar' <> 'bär'.

Natural language collations compare strings on different levels:
- 'bar' and 'bär' are identical on the first level (base character)
- 'bar' and 'bär' are different on the second level (accent)
- there are two more levels, the third being case

Strings are ordered by the first level first, then by the second, and so on.

I recommend reading Peter's excellent blog:
http://peter.eisentraut.org/blog/2023/05/16/overview-of-icu-collation-settings

So you end up with 'bar' < 'bär' < 'bat', because the first two compare
equal on level 1.

What you are looking for is a collation where accents are a first-level
difference. The only way to do that with ICU collations, as far as I know,
is to add explicit rules, like in this example:
https://stackoverflow.com/a/77288282/6464308

> Also, is there a way to see what options are there for the default built-in collations?
> I don't see, for example, the used "ks" level in the "pg_collation" table data.

You can see that in the "colllocale" column. The name of the ICU locale
determines its capabilities.

Yours,
Laurenz Albe