discourse/app/services/search_indexer.rb
Sam bd32912c5e
FIX: do not allow title stuffing to dominate search (#21464)
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.
2023-05-10 11:47:58 +10:00

419 lines
12 KiB
Ruby

# frozen_string_literal: true
class SearchIndexer
MIN_POST_BLURB_INDEX_VERSION = 4
POST_INDEX_VERSION = 5
TOPIC_INDEX_VERSION = 4
CATEGORY_INDEX_VERSION = 3
USER_INDEX_VERSION = 3
TAG_INDEX_VERSION = 3
# version to apply when issuing a background reindex
REINDEX_VERSION = 0
TS_VECTOR_PARSE_REGEX = /('([^']*|'')*'\:)(([0-9]+[A-D]?,?)+)/
def self.disable
@disabled = true
end
def self.enable
@disabled = false
end
def self.update_index(table:, id:, a_weight: nil, b_weight: nil, c_weight: nil, d_weight: nil)
raw_data = { a: a_weight, b: b_weight, c: c_weight, d: d_weight }
# The version used in excerpts
search_data = raw_data.transform_values { |data| Search.prepare_data(data || "", :index) }
# The version used to build the index
indexed_data =
search_data.transform_values do |data|
data.gsub(/\S+/) { |word| word[0...SiteSetting.search_max_indexed_word_length] }
end
table_name = "#{table}_search_data"
foreign_key = "#{table}_id"
# for user login and name use "simple" lowercase stemmer
stemmer = table == "user" ? "simple" : Search.ts_config
ranked_index = <<~SQL
setweight(to_tsvector('#{stemmer}', #{Search.wrap_unaccent("coalesce(:a,''))")}, 'A') ||
setweight(to_tsvector('#{stemmer}', #{Search.wrap_unaccent("coalesce(:b,''))")}, 'B') ||
setweight(to_tsvector('#{stemmer}', #{Search.wrap_unaccent("coalesce(:c,''))")}, 'C') ||
setweight(to_tsvector('#{stemmer}', #{Search.wrap_unaccent("coalesce(:d,''))")}, 'D')
SQL
tsvector = DB.query_single("SELECT #{ranked_index}", indexed_data)[0]
additional_lexemes = []
# we also want to index parts of a domain name
# that way stemmed single word searches will match
additional_words = []
tsvector
.scan(/'(([a-zA-Z0-9]+\.)+[a-zA-Z0-9]+)'\:([\w+,]+)/)
.reduce(additional_lexemes) do |array, (lexeme, _, positions)|
count = 0
if lexeme !~ /\A(\d+\.)?(\d+\.)*(\*|\d+)\z/
loop do
count += 1
break if count >= 10 # Safeguard here to prevent infinite loop when a term has many dots
term, _, remaining = lexeme.partition(".")
break if remaining.blank?
additional_words << [term, positions]
array << "'#{remaining}':#{positions}"
lexeme = remaining
end
end
array
end
extra_domain_word_terms =
if additional_words.length > 0
DB
.query_single(
"SELECT to_tsvector(?, ?)",
stemmer,
additional_words.map { |term, _| term }.join(" "),
)
.first
.scan(TS_VECTOR_PARSE_REGEX)
.map do |term, _, indexes|
new_indexes =
indexes
.split(",")
.map do |index|
existing_positions = additional_words[index.to_i - 1]
if existing_positions
existing_positions[1]
else
index
end
end
.join(",")
"#{term}#{new_indexes}"
end
.join(" ")
end
tsvector = "#{tsvector} #{additional_lexemes.join(" ")} #{extra_domain_word_terms}"
if (max_dupes = SiteSetting.max_duplicate_search_index_terms) > 0
reduced = []
tsvector
.scan(TS_VECTOR_PARSE_REGEX)
.each do |term, _, indexes|
family_counts = Hash.new(0)
new_index_array = []
indexes
.split(",")
.each do |index|
family = nil
family = index[-1] if index[-1].match?(/[A-D]/)
# title dupes can completely dominate the index
# so we limit them to 1
if (family_counts[family] += 1) <= (family == "A" ? 1 : max_dupes)
new_index_array << index
end
end
reduced << "#{term.strip}#{new_index_array.join(",")}"
end
tsvector = reduced.join(" ")
end
indexed_data =
if table.to_s == "post"
clean_post_raw_data!(search_data[:d])
else
search_data.values.select { |d| d.length > 0 }.join(" ")
end
params = {
"raw_data" => indexed_data,
"#{foreign_key}" => id,
"locale" => SiteSetting.default_locale,
"version" => const_get("#{table.upcase}_INDEX_VERSION"),
"search_data" => tsvector,
}
yield params if block_given?
table_name.camelize.constantize.upsert(params)
rescue => e
if Rails.env.test?
raise
else
# TODO is there any way we can safely avoid this?
# best way is probably pushing search indexer into a dedicated process so it no longer happens on save
# instead in the post processor
Discourse.warn_exception(
e,
message: "Unexpected error while indexing #{table} for search",
env: {
id: id,
},
)
end
end
def self.update_topics_index(topic_id, title, cooked)
# a bit inconsistent that we use title as A and body as B when in
# the post index body is D
update_index(
table: "topic",
id: topic_id,
a_weight: title,
b_weight: HtmlScrubber.scrub(cooked)[0...Topic::MAX_SIMILAR_BODY_LENGTH],
)
end
def self.update_posts_index(
post_id:,
topic_title:,
category_name:,
topic_tags:,
cooked:,
private_message:
)
update_index(
table: "post",
id: post_id,
a_weight: topic_title,
b_weight: category_name,
c_weight: topic_tags,
# The tsvector resulted from parsing a string can be double the size of
# the original string. Since there is no way to estimate the length of
# the expected tsvector, we limit the input to ~50% of the maximum
# length of a tsvector (1_048_576 bytes).
d_weight: HtmlScrubber.scrub(cooked)[0..600_000],
) { |params| params["private_message"] = private_message }
end
def self.update_users_index(user_id, username, name, custom_fields)
update_index(
table: "user",
id: user_id,
a_weight: username,
b_weight: name,
c_weight: custom_fields,
)
end
def self.update_categories_index(category_id, name)
update_index(table: "category", id: category_id, a_weight: name)
end
def self.update_tags_index(tag_id, name)
update_index(table: "tag", id: tag_id, a_weight: name.downcase)
end
def self.queue_category_posts_reindex(category_id)
return if @disabled
DB.exec(<<~SQL, category_id: category_id, version: REINDEX_VERSION)
UPDATE post_search_data
SET version = :version
FROM posts
INNER JOIN topics ON posts.topic_id = topics.id
INNER JOIN categories ON topics.category_id = categories.id
WHERE post_search_data.post_id = posts.id
AND categories.id = :category_id
SQL
end
def self.queue_users_reindex(user_ids)
return if @disabled
DB.exec(<<~SQL, user_ids: user_ids, version: REINDEX_VERSION)
UPDATE user_search_data
SET version = :version
WHERE user_search_data.user_id IN (:user_ids)
SQL
end
def self.queue_post_reindex(topic_id)
return if @disabled
DB.exec(<<~SQL, topic_id: topic_id, version: REINDEX_VERSION)
UPDATE post_search_data
SET version = :version
FROM posts
WHERE post_search_data.post_id = posts.id
AND posts.topic_id = :topic_id
SQL
end
def self.index(obj, force: false)
return if @disabled
category_name = nil
tag_names = nil
topic = nil
if Topic === obj
topic = obj
elsif Post === obj
topic = obj.topic
end
category_name = topic.category&.name if topic
if topic
tags = topic.tags.select(:id, :name).to_a
if tags.present?
tag_names =
(tags.map(&:name) + Tag.where(target_tag_id: tags.map(&:id)).pluck(:name)).join(" ")
end
end
if Post === obj && obj.raw.present? &&
(force || obj.saved_change_to_cooked? || obj.saved_change_to_topic_id?)
if topic
SearchIndexer.update_posts_index(
post_id: obj.id,
topic_title: topic.title,
category_name: category_name,
topic_tags: tag_names,
cooked: obj.cooked,
private_message: topic.private_message?,
)
SearchIndexer.update_topics_index(topic.id, topic.title, obj.cooked) if obj.is_first_post?
end
end
if User === obj && (obj.saved_change_to_username? || obj.saved_change_to_name? || force)
SearchIndexer.update_users_index(
obj.id,
obj.username_lower || "",
obj.name ? obj.name.downcase : "",
obj.user_custom_fields.searchable.map(&:value).join(" "),
)
end
if Topic === obj && (obj.saved_change_to_title? || force)
if obj.posts
if post = obj.posts.find_by(post_number: 1)
SearchIndexer.update_posts_index(
post_id: post.id,
topic_title: obj.title,
category_name: category_name,
topic_tags: tag_names,
cooked: post.cooked,
private_message: obj.private_message?,
)
SearchIndexer.update_topics_index(obj.id, obj.title, post.cooked)
end
end
end
if Category === obj && (obj.saved_change_to_name? || force)
SearchIndexer.queue_category_posts_reindex(obj.id)
SearchIndexer.update_categories_index(obj.id, obj.name)
end
if Tag === obj && (obj.saved_change_to_name? || force)
SearchIndexer.update_tags_index(obj.id, obj.name)
end
end
def self.clean_post_raw_data!(raw_data)
urls = Set.new
raw_data.scan(Discourse::Utils::URI_REGEXP) { urls << $& }
urls.each do |url|
begin
case File.extname(URI(url).path || "")
when Oneboxer::VIDEO_REGEX
raw_data.gsub!(url, I18n.t("search.video"))
when Oneboxer::AUDIO_REGEX
raw_data.gsub!(url, I18n.t("search.audio"))
end
rescue URI::InvalidURIError
end
end
raw_data
end
private_class_method :clean_post_raw_data!
class HtmlScrubber < Nokogiri::XML::SAX::Document
attr_reader :scrubbed
def initialize
@scrubbed = +""
end
def self.scrub(html)
return +"" if html.blank?
begin
document = Nokogiri.HTML5("<div>#{html}</div>", nil, Encoding::UTF_8.to_s)
rescue ArgumentError
return +""
end
nodes = document.css("div.#{CookedPostProcessor::LIGHTBOX_WRAPPER_CSS_CLASS}")
if nodes.present?
nodes.each do |node|
node.traverse do |child_node|
next if child_node == node
if %w[a img].exclude?(child_node.name)
child_node.remove
elsif child_node.name == "a"
ATTRIBUTES.each { |attribute| child_node.remove_attribute(attribute) }
end
end
end
end
document.css("img.emoji").each { |node| node.remove_attribute("alt") }
document
.css("a[href]")
.each do |node|
if node["href"] == node.text || MENTION_CLASSES.include?(node["class"])
node.remove_attribute("href")
end
if node["class"] == "anchor" && node["href"].starts_with?("#")
node.remove_attribute("href")
end
end
html_scrubber = new
Nokogiri::HTML::SAX::Parser.new(html_scrubber).parse(document.to_html)
html_scrubber.scrubbed.squish
end
MENTION_CLASSES ||= %w[mention mention-group]
ATTRIBUTES ||= %w[alt title href data-video-title]
def start_element(_name, attributes = [])
attributes = Hash[*attributes.flatten]
ATTRIBUTES.each do |attribute_name|
if attributes[attribute_name].present? &&
!(attribute_name == "href" && UrlHelper.is_local(attributes[attribute_name]))
characters(attributes[attribute_name])
end
end
end
def characters(str)
scrubbed << " #{str} "
end
end
end