Fuzzy matching conundrum

[TheLeoP]

The problem

I encountered this issue while using emmet-language-server. If I write

ul>li.item$*

with defaults settings, I get an snippet completion candidate that expands the snippet. However, if I write

ul>li.item$*5

I don't get anything.

I dug into the code and play with the configuration and found that the reason for this was fuzzy_cutoff (unifying_chars could technically also help, but keep reading), if I set it to 0, I get the expected completion result in both cases. This happens because

coq_nvim/coq/clients/lsp/worker.py

Lines 43 to 48 in 976012b

	cword = cword_before(
	match.unifying_chars,
	lower=True,
	context=context,
	sort_by=sort_by,
	)

In the first case, cword is an empty string. Since * is not isalnum and isn't in an unifyin_char (from is_word), the fuzzy matching is performed against context.l_words_before. Again, since * is a symbol because of is_word, context.l_words_before is empty. This makes ratio 1 and is passes the default fuzzy_cutoff value

coq_nvim/coq/clients/lsp/worker.py

Lines 51 to 55 in 976012b

	ratio = multi_set_ratio(
	cword,
	lower(sort_by),
	look_ahead=match.look_ahead,
	)

In the second case, cword is 5 because 5 is isalnum. This makes ratio 0, so it doesn't pass the default fuzzy_cutoff value.

unyfing_chars

"But, that's what unyfing_chars is for, right?" I thought. For most symbols, yes. If I add all of the symbols used in the emmet syntax to unyfing_chars I always get the expected snippet completion. The problem is that said symbols contain things like ., which isn't a problem for html, but it is for other languages.

For example, lua_ls (when configured to use the Neovim runtime libraries) gives two completion candidates for

vim.keymap.

both are snippets for the functions set and del.

With the default values for unyfing_chars, set and del are matched against an empty string (because . is a symbol, but vim.keymap.set starts with a char that is is_word, so it returns context.l_words_before, which is empty), so ratio is 1. They are shown.

With the modified values for unyfing_chars, set and del are matched against vim.keymap, the ratio is 0. They aren't shown.

Proposed solutions

filetype specific configurations

Cons:

• Could get cumbersome on the plugin side
• Still won't work for tsx/jsx unless some kind of autocmd changes the configuration dinamically depending on the context on the user side

modify what the fuzzy matching is made against

Cons:

• I can't think of a better alternative that won't also have its own limitations. In general, I thinks this is a hard problem without some kind of per-language configuration

allow manual triggered to (maybe optionally) skip fuzzy_cutoff check

Changing

coq_nvim/coq/clients/lsp/worker.py

Lines 56 to 61 in 976012b

	use = ratio >= match.fuzzy_cutoff and (
	isinstance(comp.primary_edit, SnippetEdit)
	or bool(comp.secondary_edits)
	or bool(comp.extern)
	or not cword.startswith(comp.primary_edit.new_text)
	)

to

        use = (
            ratio >= match.fuzzy_cutoff
            and (
                isinstance(comp.primary_edit, SnippetEdit)
                or bool(comp.secondary_edits)
                or bool(comp.extern)
                or not cword.startswith(comp.primary_edit.new_text)
            )
            or context.manual
        )

would allow users to skip this kind of check manually.

possible workarounds

Using something like

ul>li.item$*5.

doesn't changes the result from the language server and (incidentally) makes coq show the completion because . makes it match against an empty string (so ratio is 1)