We were giving topics with repeated words extra weight in search index.
This meant that it was trivial to stuff words into title to dominate in search
given we search for exact title matches first.
The following tweak means that:
`invite invited invites`
and
`invite some stuff`
Both rank the same for title searching.
Titles are short and punchy, duplicating words should not give special
weight.
Requires a full reindex to take effect.
During search indexing we "stuff" the index with additional keywords for
entities that look like domain names.
This allows searches for `cnn` to find URLs for `www.cnn.com`
The search stuffing attempted to keep indexes aligned at the correct positions
by remapping the indexed terms. However under certain edge cases a single
word can stem into 2 different lexemes. If this happened we had an off by
one which caused the entire indexing to fail.
We work around this edge case (and carry incorrect index positions) for cases
like this. It is unlikely to impact search quality at all given index position
makes almost no difference in the search algorithm.
- Reduce duplication of terms in post index from unlimited to 6. This will
result in reduced index size and reduced weighting for posts containing
a huge amount of duplicate terms. (Eg: a post containing "sam sam sam sam
sam sam sam sam", will index as "sam sam sam sam sam sam", only including
the word up to 6 times.) This corrects a flaw where title weighting could
be ignored.
- Prioritize exact matches of words in titles. Our search always performs
a prefix match. However we want to give special weight to exact title matches
meaning that a search for "sum" will find topics such as "the sum of us" vs
"summer in spring".
- Pick up fixes to our search algorithm which are missing from old indexes.
Specifically pick up the fix that indexes URLs properly. (`https://happy.com`
was stemmed to `happi` in keywords and then was not searchable)
see also:
https://meta.discourse.org/t/refinements-to-search-being-tested-on-meta/254158
Indexing will take a while and work in batches, in the background.
If a post contains domain with a word that stems to a non prefix single
words will not match it.
For example: in happy.com, `happy` stems to `happi`. Thus searches for happy
will not find URLs with it included.
This bloats the index a tiny bit, but impact is limited.
Will require a full reindex of search to take effect.
When we are done refining search we can consider a full version bump.
Previous regex did not allow for cases where a lexeme contains a : (colon)
This can happen when parsing URLs. New algorithm allows for this.
Test was amended to more clearly call out index problems
* FEATURE: allow restricting duplication in search index
This introduces the site setting `max_duplicate_search_index_terms`.
Using this number we limit the amount of duplication in our search index.
This allows us to more correctly weight title searches, so bloated posts
don't unfairly bump to the top of search results.
This feature is completely disabled by default and behind a site setting
We will experiment with it first. Note entire search index must be rebuilt
for it to take effect.
---------
Co-authored-by: Alan Guo Xiang Tan <gxtan1990@gmail.com>
This makes it easier to find PMs involving a particular user, for
example by searching for `in:messages thisUser` (previously, that query
would only return results in posts where `thisUser` was in the post body).
The search_ignore_accents site setting can be used to make the search
indexer remove the accents before indexing the content. The unaccent
function from PostgreSQL is better than Ruby's unicode_normalize(:nfkd).
Random strings can result into much longer tsvectors. For example
parsing a Base64 string of ~600kb can result in a tsvector of over 1MB,
which is the maximum size of a tsvector.
Follow-up-to: 823c3f09d4
This commit fixes a bug where we our `HTMLScrubber` was only searching
for emoji img tags which contains only the "emoji" class. However, our emoji image tags
may contain more than just the "emoji" class like "only-emoji" when an
emoji exists by itself on a single line.
Long posts may have `cooked` fields that produce tsvectors longer than
the maximum size of 1MiB (1,048,576 bytes). This commit uses just the
first million characters of the scrubbed cooked text for indexing.
Reducing the size to exactly 1MB (1_048_576) is not sufficient because
sometimes the output tsvector may be longer than the input and this
gives us some breathing room.
Over the years we accrued many spelling mistakes in the code base.
This PR attempts to fix spelling mistakes and typos in all areas of the code that are extremely safe to change
- comments
- test descriptions
- other low risk areas
When the admin creates a new custom field they can specify if that field should be searchable or not.
That setting is taken into consideration for quick search results.
In the near future, we will be swtiching to PG headlines to generate the
search blurb. As such, we need to replace audio and video links in the
raw data used for headline generation. This also means that we avoid
replacing links each time we need to generate the blurb.
We do prefix matching in search so there is no need to inject the extra
terms.
Before:
```
"'discourse':10,11 'discourse.org':10,11 'org':10,11 'test':8A,10,11 'test.discourse.org':10,11 'titl':4A 'uncategor':9B"
```
After:
```
"'discourse.org':10,11 'org':10,11 'test':8A 'test.discourse.org':10,11 'titl':4A 'uncategor':9B"
```
```
discourse_development=# SELECT alias, lexemes FROM TS_DEBUG('www.discourse.org');
alias | lexemes
-------+---------------------
host | {www.discourse.org}
discourse_development=# SELECT TO_TSVECTOR('www.discourse.org');
to_tsvector
-----------------------
'www.discourse.org':1
```
Given the above lexeme, we will inject additional lexeme by splitting
the host on `.`. The actual tsvector stored will look something like
```
tsvector
---------------------------------------
'discourse':1 'discourse.org':1 'org':1 'www':1 'www.discourse.org':1
```
There is a feature in search where we take over from the tokenizer
in postgres and attempt to inject more words into search.
So for example: sam.i.am will inject the words i and am.
This is not ideal cause there are many edge cases and this can
cause extreme index bloat.
This is an opening move commit to make it configurable, over the
next few weeks we will evaluate and decide if we disable this by
default or simply remove.
This feature adds the ability to define synonyms for tags, and the ability to merge one tag into another while keeping it as a synonym. For example, tags named "js" and "java-script" can be synonyms of "javascript". When searching and creating topics using synonyms, they will be mapped to the base tag.
Along with this change is a new UI found on each tag's page (for example, `/tags/javascript`) where more information about the tag can be shown. It will list the synonyms, which categories it's restricted to (if any), and which tag groups it belongs to (if tag group names are public on the `/tags` page by enabling the "tags listed by group" setting). Staff users will be able to manage tags in this UI, merge tags, and add/remove synonyms.
Zeitwerk simplifies working with dependencies in dev and makes it easier reloading class chains.
We no longer need to use Rails "require_dependency" anywhere and instead can just use standard
Ruby patterns to require files.
This is a far reaching change and we expect some followups here.
Previous to this fix is a post had the test www.test.com/abc it would fail
to index.
This also simplifies the rules to avoid full url parsing which can be
expensive
This commit fixes the follow quality issue with `PostSearchData#raw_data`:
1. URLs are being tokenized and links with similar href and characters
are being duplicated in the raw data.
`Post#cooked`:
```
<p><a href=\"https://meta.discourse.org/some.png\" class=\"onebox\" target=\"_blank\" rel=\"nofollow noopener\">https://meta.discourse.org/some.png</a></p>
```
`PostSearchData#raw_data` Before:
```
This is a test topic 0 Uncategorized https://meta.discourse.org/some.png discourse org/some png https://meta.discourse.org/some.png discourse org/some png
```
`PostSearchData#raw_data` After:
```
This is a test topic 0 Uncategorized https://meta.discourse.org/some.png meta discourse org
```
2. Ligthbox being included in search pollutes the
`PostSearchData#raw_data` unncessarily.
From 28 March 2018 to 28 March 2019, searches for the term `image` on
`meta.discourse.org` had a click through rate of 2.1%. Non-lightboxed images are not included in indexing for search yet we were indexing content within a lightbox. Also, search for terms like `image` was affected we were using `Pasted image` as the filename for
uploads that were pasted.
`Post#cooked`
```
<p>Let me see how I can fix this image<br>\n<div class=\"lightbox-wrapper\"><a class=\"lightbox\" href=\"https://meta.discourse.org/some.png\" title=\"some.png\" rel=\"nofollow noopener\"><img src=\"https://meta.discourse.org/some.png\" width=\"275\" height=\"299\"><div class=\"meta\">\n<svg class=\"fa d-icon d-icon-far-image svg-icon\" aria-hidden=\"true\"><use xlink:href=\"#far-image\"></use></svg><span class=\"filename\">some.png</span><span class=\"informations\">1750×2000</span><svg class=\"fa d-icon d-icon-discourse-expand svg-icon\" aria-hidden=\"true\"><use xlink:href=\"#discourse-expand\"></use></svg>\n</div></a></div></p>
```
`PostSearchData#raw_data` Before:
```
This is a test topic 0 Uncategorized Let me see how I can fix this image some.png png https://meta.discourse.org/some.png discourse org/some png some.png png 1750×2000
```
`PostSearchData#raw_data` After:
```
This is a test topic 0 Uncategorized Let me see how I can fix this image
```
In terms of indexing performance, we now have to parse the given HTML
through nokogiri twice. However performance is not a huge worry here since a string length of 194170 takes only 30ms
to scrub plus the indexing takes place in a background job.
* This is causing certain posts to appear in searches incorrectly as `PostSearchData#raw_data` contains the outdated title, category name and tag names.
4481836 introduced accent stipping in search_indexer,
but we need to strip it from the query itself as well
TODO in search with diacritics:
- Still need to fix excerpts on search page
- need to support accent stripping in in_topic search
- need to make sure that in:title works correctly
- need to fix "word boldening" in titles
Sam
Jan Cernik
Daniel Waterworth
David Taylor
Penar Musaraj
Bianca Nenciu
Dan Ungureanu
Alan Guo Xiang Tan
Josh Soref
Krzysztof Kotlarek
Neil Lalonde
Gerhard Schlager
Guo Xiang Tan
Daniel Hollas
