Re: package-autosuggest

[Philip Kaludercic]

Eshel Yaron <me@eshelyaron.com> writes:

> Hello Stefan,
>
> Stefan Kangas <stefankangas@gmail.com> writes:
>
>> Eshel Yaron <me@eshelyaron.com> writes:
>>
>>> That'd indeed already be an improvement, my point is that in some cases
>>> we would know that it really is the right package with very high
>>> certainty.
>>
>> Shouldn't that just be the criterion, then?  In other words, isn't
>> that what it would mean to only recommend things that are likely to be
>> helpful?
>
> I'm not sure.  In Philip's draft, I don't think that this is the case.
> (Correct me if I've missed anything, please.)  

The assumption in my draft was that the database would only be
populated with entries that would be considered helpful, or rather if
the database were to be generated from GNU ELPA and NonGNU ELPA
packages, it shouldn't recommend unhelpful packages, as all packages in
GNU and NonGNU ELPA are supposed to be helpful in the first place.

>                                                The way I see it there
> are two parameters to consider for each recommendation.  Let's call them
> potential and confidence.  Potential is how much value the user can
> obtain from Emacs's recommendation.  That depends mostly on the
> recommended package itself.  Confidence is how certain Emacs is that
> this user should use this package.  For example, again, in the current
> draft we have a recommendation for `sml-mode` based on files with
> extension `.sml`.  This recommendation has great potential to benefit
> Standard ML users (I assume, I haven't tried it out myself), but the
> confidence for this recommendation isn't that high, because it's prone
> to false positives (not all `.sml` files are Standard ML).

That is a legitimate point, but one that Emacs suffers from in general.
I frequently find Perl and Prolog files that both end in .pl, but there
can only be one entry in auto-mode-list (unless you manually do a
`c-or-c++-mode').

> The crux is that the quality of a recommendation depends not only on the
> quality of the package but also on the strength of the signal that leads
> Emacs to recommend it (the quality of the "hint").

So are you thinking of a numerical value?  And wouldn't the value depend
on the file contents?  And how do you think this valuation should be
used?

> Personally, I think that even with a chance of false positives,
> suggesting `sml-mode` to a user that opens `foo.sml` is great.  For this
> reason, I proposed that Emacs should make both high-confidence and
> low-confidence recommendations, but use different messages for the two
> cases (or three, if you want to also have "medium-confidence").

The current proposal wouldn't install anything automatically anyway, it
would prompt the user if they are interested in this kind of a file

>
> Best,
>
> Eshel