Hi Simon,

On 9/13/24 10:12, Simon Tournier wrote:

>> | tor                   | Tor related; ~torbrowser~ somewhere near top. |
> --8<---------------cut here---------------start------------->8---
> $ ./pre-inst-env guix search tor | recsel -p name,relevance | head -8
> name: tor
> relevance: 208
>
> name: tor-client
> relevance: 169
>
> name: torsocks
> relevance: 103
> --8<---------------cut here---------------end--------------->8---
>
> Compared to current:
>
> --8<---------------cut here---------------start------------->8---
> $ guix search tor | recsel -p name,relevance | head -8
> name: tor
> relevance: 47
>
> name: ghc-storablevector
> relevance: 29
>
> name: tor-client
> relevance: 28
> --8<---------------cut here---------------end--------------->8---
>
> However, the position move from 225th to 19th.
>
>      $ guix search tor | recsel -P name | grep -n torbrowser
>      225:torbrowser
>
>      $ ./pre-inst-env guix search tor | recsel -P name | grep -n torbrowser
>      19:torbrowser
>
> Similarly as ’dig’, the description of ’torbrowser’ package could be
> improvement.  Because ’guix search tor browser’ returns nothing.

Does ~torbrowser~ not appear as the third result in all three cases for 
you when running =guix search tor browser=?

Otherwise, if you meant =guix search tor= to find ~torbrowser~: perhaps 
it should be higher ranked, but it could be argued that patch v1's 
behavior is still more optimal in this aspect considering all results 
above ~torbrowser~ it are indeed related to Tor.

>> | Keyword(s) with poor  | Expectations                                  |
>> | results before        |                                               |
>> |-----------------------+-----------------------------------------------|
>> | dig                   | ~bind~ near top.                              |
> Hum, indeed and I do not know if we can improve here.  Well, it’s hard
> to improve for short terms, BTW.
>
> --8<---------------cut here---------------start------------->8---
> $ ./pre-inst-env guix search dig | recsel -p name,relevance | head -8
> name: go-go-uber-org-dig
> relevance: 104
>
> name: rust-num-bigint-dig
> relevance: 78
>
> name: rust-num-bigint-dig
> relevance: 78
> --8<---------------cut here---------------end--------------->8---
>
> Compared to current:
>
> --8<---------------cut here---------------start------------->8---
> $ guix search dig | recsel -p name,relevance | head -8
> name: sysdig
> relevance: 24
>
> name: texlive-pedigree-perl
> relevance: 13
>
> name: ruby-net-http-digest-auth
> relevance: 13
> --8<---------------cut here---------------end--------------->8---
>
> Indeed, 17th position is better than 609th.  But if you add a term as
> ’dns’, bang! :-)  Well, BTW the description of ’bind’ could be a bit
> improved because the word network does not appear.  Anyway. :-)

[...]

>> | rsh                   | ~inetutils~ near top.                         |
> --8<---------------cut here---------------start------------->8---
> $ ./pre-inst-env guix search rsh | recsel -p name,relevance | head -8
> name: inetutils
> relevance: 26
>
> name: emacs-tramp
> relevance: 26
>
> name: rust-borsh-schema-derive-internal
> relevance: 22
> --8<---------------cut here---------------end--------------->8---
>
> Compared to current:
>
> --8<---------------cut here---------------start------------->8---
> $ guix search rsh | recsel -p name,relevance | head -8
> name: go-sigs-k8s-io-yaml
> relevance: 14
>
> name: python-pymarshal
> relevance: 13
>
> name: emacs-powershell
> relevance: 13
> --8<---------------cut here---------------end--------------->8---

[...]

>> | gcc                   | ~gcc-toolchain~ near top.                     |
> Indeed, something is unexpected.  Well, first:
>
>      $ guix search gcc | recsel -CP name | uniq | head -8
>      gccgo
>      gfortran-toolchain
>      gdc-toolchain
>      gcc-toolchain
>      gcc-cross-x86_64-w64-mingw32-toolchain
>      gcc-cross-or1k-elf-toolchain
>      gcc-cross-i686-w64-mingw32-toolchain
>      gcc-cross-avr-toolchain
>
>      $ guix search gcc | recsel -CP name | uniq -c | sort -rn | head -8
>           18 llvm
>           12 gcc-toolchain
>            6 libgccjit
>            6 gccgo
>            3 isl
>            2 libstdc++-doc
>            2 java-commons-cli
>            2 gdc-toolchain
>
> Other said, the packages with multi-versions decrease the experience.
> Well, that had already by “improved” [1] with some SEO. ;-)  Indeed,
> maybe the relevance should be improved.
>
> Second, gccgo has a relevance score of 22 with the only term ’gcc’,
> compared to gcc-toolchain scoring at 15.
>
>      gccgo        gcc-toolchain
>    4 * 1 * 1      4 * 1 * 1
> + 2 * 5 * 1    + 2 * 1 * 1
> + 1 * 0        + 1 * 0
> + 3 * 1 * 1    + 3 * 1 * 1
> + 2 * 0        + 2 * 1 * 3
> + 1 * 5 * 1    + 1 * 0
> = 22           = 15
>
> This is unexpected.  And, IMHO that’s bug!  In the description of
> gcc-toolchain, the term ’gcc’ appears 3 times but it only score with ’1’
> instead of ’5’.
>
> As the patch try to address, the main issue is:
>
>    (define (score regexp str)
>      (fold-matches regexp str 0
>                    (lambda (m score)
>                      (+ score
>                         (if (string=? (match:substring m) str)
>                             5             ;exact match
>                             1)))))
>
> Here the exact match does not consider a substring exact match.  For
> instance, one would consider that the term ’gcc’ exactly matches in
> “some GCC thing”.  Considering the current implementation, that’s not
> the case.  For instance, a snippet as the procedure ’scoring’:
>
> --8<---------------cut here---------------start------------->8---
> scheme@(guix-user)> ,use(ice-9 regex)
> scheme@(guix-user)> (define regexp (make-regexp "gcc" regexp/icase))
> scheme@(guix-user)> (define str "some GCC thing")
> scheme@(guix-user)> (fold-matches regexp str 0
>      (lambda (m res)
>        (+ res
>          (if (string=? (match:substring m) str)
>            5 1))))
> $2 = 1
> --8<---------------cut here---------------end--------------->8---
>
>
> See v2 for my proposal fixing this.
>
> Please note that this v2 gives the same ranking for torbrowser.  And
> also improve the situation with gcc-toolchain.
>
> --8<---------------cut here---------------start------------->8---
> $ ./pre-inst-env guix search gcc | recsel -CP name | grep -n gcc-toolchain
> 1:gcc-toolchain
> 2:gcc-toolchain
> 3:gcc-toolchain
> 4:gcc-toolchain
> 5:gcc-toolchain
> 6:gcc-toolchain
> 7:gcc-toolchain
> 8:gcc-toolchain
> 9:gcc-toolchain
> 10:gcc-toolchain
> 11:gcc-toolchain
> 12:gcc-toolchain
>
> $ ./pre-inst-env guix search tor | recsel -CP name | grep -n torbrowser
> 7:torbrowser
>
> $ ./pre-inst-env guix search dig | recsel -CP name | grep -n bind
> 44:bind
> --8<---------------cut here---------------end--------------->8---
>
> However, inetutils is still at 44th with the only one term ’rsh’.  I
> would suggest to do some tweak with the description.

And including a relevant part of your message from #70689:

> Again, considering the case at hand: If instead of 3 randomly picked in
> v2 of #73220, we would pick 7, then inetutils is ranked first.
>
> Yeah, maybe 3 isn’t enough… And maybe 7 is a good choice.
What do you think about setting the value to the sum of all weights in 
~metrics~ as I did in patch v1? My logic is that an object is almost 
always going to be relevant if it contains a whole word match compared 
to "maybe relevant" if it only matches substrings, so it would be 
reasonable to thus show most of the objects with whole word matches 
first. This improves or maintains consistency of relevant results in the 
test cases with shorter terms, and also reduces the need for guesswork 
with choosing arbitrary numbers that may or may not work.

Note that I also gave the same treatment to exact match scores, although 
not as extremely weighed (only double the whole word score in v1).

In the case of ~inetutils~, for example, this formula guarantees that if 
I were to search =rsh= - which is a common subword, but itself has a 
very unique meaning - ~inetutils~ /always/ shows up at or near the top 
along with other rsh-related packages, assuming no exact matches.

In other words, the intention would be to have the calculations set up 
such that they implicitly "categorize" object rankings into a (rough) 
hierarchy of the following:

|--------------------------------------------| | Objects with at least 
one exact match | |--------------------------------------------| | 
Objects with at least one whole word match | 
|--------------------------------------------| | Objects with only 
substring matches | |--------------------------------------------|

>> I opted to switch to counting a maximum of one match per field, which helps
>> with cases where a common subword matches /many/ times in packages with longer
>> descriptions, pushing more relevant packages down.  In multi-term searches,
>> the unique terms - which are naturally rarer - also contribute to a larger
>> percentage of the score as a result of these changes.
>> Having matches with only one word boundary be scored as 2 instead of 1 was
>> done with the reasoning that a term is more likely to be part of a compound
>> word name (and thus more relevant) if it is a prefix or suffix; for example,
>> "gl" in OpenGL, "borg" in borgmatic, and "tor" in torbrowser.
> [...]
>
>> Closing this message on an unrelated note for future work: I stumbled on an
>> interesting idea while looking for test cases which suggested reducing the
>> score of a programming library when its language is not included in search
>> terms.  It's out of scope for the current issue, but I thought I'd mention it
>> anyways for potential further improvements.
> Well, years ago I thought about implementing TF-IDF [2,3].  Other ideas
> [4] are floating around.  Then, we spent some time for making “guix
> search” faster [5] and today my TODO is about having an extension
> relying on Guile-Xapian.
>
> Therefore, I would prefer keep the ’relevance’ more or less predictable
> by only counting the number of occurrences and apply some weights.
> Else, for what my opinion is worth, the direction would not be to
> re-invent an algorithm but maybe implement some already well-known ones.
> TF-IDF [3] is one or Okapi-BM25 is another one, etc.  In all in all,
> that what Xapian provides. ;-) And it does it very well!  That’s why I
> would be tempted to have a Guix extension relying on Guile-Xapin for
> indexing and searching (fast!).

Yes, I had thought about trying something like TF-IDF while looking into 
the issue, but it seemed much less trivial than changes to a scoring 
function. The count-once-per-field change was supposed to at least 
tangentially mimic this behavior and reduce bias towards objects that 
happen to have very long descriptions but aren't very relevant. It's 
also needed for my "categorization" math to hold.

> Hum, why this:
>
>      guix search ' dig$' dig | recsel -p name,relevance | head -8
>
> does not return the package ’bind’?

It appears the ~regexp/newline~ flag needs to be set for ~make-regexp~. 
A quick test adding it here [1] seemed to work.


My main concern with v2 is that I don't think whole words are weighed 
heavily enough, but it provides a simpler solution that still offers 
improvement, so I'm happy either way.

Thanks for the feedback!

[1] 
https://git.savannah.gnu.org/cgit/guix.git/tree/guix/scripts/package.scm#n897

Cheers,

aurtzy