How does search on AMO work?
High-level overview
AMO add-ons are indexed in our Elasticsearch cluster. For each search query someone makes on AMO, we run a custom set of full-text queries against that cluster.
Our autocomplete (that you can see when starting to type a few characters in the search field) uses the exact same implementation as a regular search underneath.
Rules
For each search query, we apply a number of rules that attempt to find the search terms in each add-on name, summary and description. Each rule generates a score that depends on:
The frequency of the terms in the field we’re looking at
The importance of each term in the overall index (the more common the term is across all add-ons, the less it impacts the score)
The length of the field (shorter fields give a higher score as the search term is considered more relevant if they make up a larger part of the field)
Each rule is also given a specific boost affecting its score, making matches against the add-on name more important and matches against the summary or description.
Add-on names receive special treatment: Partial or misspelled matches are accepted to some extent while exact matches receive a significantly higher score.
Scoring
Each score for each rule is combined into a final score which we modify depending on the add-on popularity on a logarithm scale. “Recommended” and “By Firefox” add-ons get an additional, significant boost to their score.
Finally, results are returned according to their score in descending order.
Technical overview
We store two kinds of data in the addons index: indexed fields that are used for search purposes, and non-indexed fields that are meant to be returned (often as-is with no transformations) by the search API (allowing us to return search results data without hitting the database). The latter is not relevant to this document.
Our search can be reached either via the API through /api/v5/addons/search/ or /api/v5/addons/autocomplete/ which are used by our frontend.
Indexing
The key fields we search against are name
, summary
and description
. Because all can be translated, we index them multiple times:
Once with the translation in the default locale of the add-on, under
{field}
, analyzed with just thesnowball
analyzer fordescription
andsummary
, and a custom analyzer forname
that applies the following filters:standard
,word_delimiter
(a custom version withpreserve_original
set totrue
),lowercase
,stop
, anddictionary_decompounder
(with a specific word list) andunique
.Once for every translation that exists for that field, using Elasticsearch language-specific analyzer if supported, under
{field}_l10n_{analyzer}
.
In addition, for the name, we also have:
For all fields described above also contains a subfield called
raw
that holds a non-analyzed variant for exact matches in the corresponding language (stored as akeyword
, with alowercase
normalizer).A
name.trigram
variant for the field in the default language, which is using a custom analyzer that depends on angram
tokenizer (withmin_gram=3
,max_gram=3
andtoken_chars=["letter", "digit"]
).
Flow of a search query through AMO
Let’s assume we search on addons-frontend (not legacy) the search query hits the API and gets handled by AddonSearchView
, which directly queries ElasticSearch and doesn’t involve the database at all.
There are a few filters that are described in the /api/v5/addons/search/ docs but most of them are not very relevant for text search queries. Examples are filters by guid, platform, category, add-on type or appversion (application version compatibility). Those filters are applied using a filter
clause and shouldn’t affect scoring.
Much more relevant for text searches (and this is primarily used when you use the search on the frontend) is SearchQueryFilter
.
It composes various rules to define a more or less usable ranking:
Primary rules
These are the ones using the strongest boosts, so they are only applied to the add-on name.
Applied rules (merged via should
):
A
dis_max
query withterm
matches onname_l10n_{analyzer}.raw
andname.raw
if the language of the request matches a known language-specific analyzer, or just aterm
query onname.raw
(boost=100.0
) otherwise - our attempt to implement exact matchesIf we have a matching language-specific analyzer, we add a
match
query toname_l10n_{analyzer}
(boost=5.0
,operator=and
)A
phrase
match onname
that allows swapped terms (boost=8.0
,slop=1
)A
match
onname
, using the standard text analyzer (boost=6.0
,analyzer=standard
,operator=and
)A
prefix
match onname
(boost=3.0
)If a query is < 20 characters long, a
dis_max
query (boost=4.0
) composed of a fuzzy match onname
(boost=4.0
,prefix_length=2
,fuzziness=AUTO
,minimum_should_match=2<2 3<-25%
) and amatch
query onname.trigram
, with aminimum_should_match=66%
to avoid noise
Secondary rules
These are the ones using the weakest boosts, they are applied to fields containing more text like description, summary and tags.
Applied rules (merged via should
):
Look for matches inside the summary (
boost=3.0
,operator=and
)Look for matches inside the description (
boost=2.0
,operator=and
)
If the language of the request matches a known language-specific analyzer, those are made using a multi_match
query using summary
or description
and the corresponding {field}_l10n_{analyzer}
, similar to how exact name matches are performed above, in order to support potential translations.
Scoring
We combine scores through a function_score
query that multiplies the score by several factors:
A first multiplier is always applied through the
field_value_factor
function onaverage_daily_users
with alog2p
modifierAn additional
4.0
weight is applied if the add-on is public & non-experimental.Finally,
5.0
weight is applied to By Firefox and Recommended add-ons.
On top of the two sets of rules above, a rescore
query is applied with a window_size
of 10
. In production, we have 5 shards, so that should re-adjust the score of the top 50 results returned only. The rules used for rescoring are the same used in the secondary rules above, with just one difference: it’s using match_phrase
instead of match
, with a slop of 10
.
General query flow
Fetch current translation
Fetch locale specific analyzer (List of analyzers)
Apply primary and secondary should rules
Determine the score
Rescore the top 10 results per shard