I have indexed all wikipedia pages on elasticsearch, and now I would like to search through them according to a list of keywords that I have created. The documents on elasticsearch have only three fields: id for the page id, title for the page title and content for the page content (already clean of wikipedia markup).
My goal is to reproduce the mediawiki query api as much as possible, with parameters action=query and list=search. For instance, given the keywords "non riemannian metric spaces", a call to
gives a list of the most relevant pages for those keywords.
So far I have been using rather simple elasticsearch search queries, like for instance
POST _search
{
"query": {
"bool" : {
"must" : {
"match" : {
"content": {
"query": "non riemannian metric spaces"
}
}
},
"should" : {
"match" : {
"title": {
"query": "non riemannian metric spaces",
"boost": x
}
}
}
}
}
}
for several values of boost, like 1, 2 or 0.5. This gives already some decent results, in the sense that the pages I obtain are relevant to the keywords, but sometimes they are not quite the same I get with the mediawiki api.
I would be glad to hear some suggestions on how to fine-tune the elasticsearch query to mimic more accurately the mediawiki api behavior. Or even, since the mediawiki api itself is built with elasticsearch and cirrussearch, I would like to know whether the actual elasticsearch query for the entry point above with those specific parameters is openly available.
Thank you in advance!
UPDATE (after Robis Koopmans' answer): Seeing the actual query with cirrusDumpQuery has indeed been very useful. I do however have some followup questions concerning the query:
The query has a set of similar multi_match clauses searching my keywords in fields like ["title.plain^1", "title^3"]. While I understand the ^n boost, I ignore what .plain refers to. Does it have to do with elasticsearch itself (i.e. is it a field derived from title at index time?) or is it something that has to do with the specific mediawiki mapping they use? In any case, I would appreciate some more information about this.
At some other point in the query, there is a {"match": {"all": {...}}} clause. What exactly is the all key here? Is it a document field? Is it related with the match_all clause?
What is the suggest field that appears in the query? In the score explanation it seems to be associated with synonyms. How are those handled in this case?
To be performed after the search, there is a rescore clause with two other score functions. One of them uses the popularity_score of a wikipedia page. What is that?
And finally, the most relevant score that ends up ranking the pages is the output of the sltr clause. In it, there is a "model": "enwiki-20220421-20180215-query_explorer", and in the score explanation it is identified with a LtrModel: naive_additive_decision_tree. I understand that this model is some stored LTR model. However, since it seems to be the most relevant number in the final sorting of the results, what exactly is that model and is it openly available?
Please feel free to answer whichever question you know the answer to, and again thanks a lot!
The query has a set of similar multi_match clauses searching my keywords in fields like ["title.plain^1", "title^3"]. While I understand the ^n boost, I ignore what .plain refers to. Does it have to do with elasticsearch itself (i.e. is it a field derived from title at index time?) or is it something that has to do with the specific mediawiki mapping they use? In any case, I would appreciate some more information about this.
The .plain fields are generated as part of the elasticsearch mapping. The current settings and mappings are available to see how exactly they work. mediawiki.org includes a search glossary entry on the plain field as well. In general the top level field contains a highly processed form of the text, and the plain field uses minimal analysis.
At some other point in the query, there is a {"match": {"all": {...}}} clause. What exactly is the all key here? Is it a document field? Is it related with the match_all clause?
mediawiki.org also contains an (incomplete) CirrusSearch schema that gives a brief description of these fields and the various analysis chain components used. The all field is an optimization to give a strong first-pass filter against the search index.
What is the suggest field that appears in the query? In the score explanation it seems to be associated with synonyms. How are those handled in this case?
Suggest field contains shingles (word ngrams) of the articles title and redirects, essentially a pre-calculation of phrase queries. The suggest might look like synonyms in the explain output, and they often contain those, but it also includes misspellings, translations, and numerous other reasons editors have for creating redirects. Matches on redirects are generally a strong relevance signal.
To be performed after the search, there is a rescore clause with two other score functions. One of them uses the popularity_score of a wikipedia page. What is that?
This is the fraction of page views on the wiki that go to that article.
And finally, the most relevant score that ends up ranking the pages is the output of the sltr clause. In it, there is a "model": "enwiki-20220421-20180215-query_explorer", and in the score explanation it is identified with a LtrModel: naive_additive_decision_tree. I understand that this model is some stored LTR model. However, since it seems to be the most relevant number in the final sorting of the results, what exactly is that model and is it openly available?
This model is generated by mjolnir and essentially overwrites the score from the rest of the query. There is some information in wikitech (found there as it is more specific to the WMF deployment of mediawiki than mediawiki itself), also a slide deck called From Clicks to Models might give some insight into whats happening in that code base. Perhaps important to know mjolnir only applies to bag of words queries, queries invoking phrases or other expert functionality skip the ML model.
Noone had asked for the models before, if they might be useful i dumped the current models from the ranking plugin. This contains both the feature definitions used and the decision trees generated by xgboost.
I didn't find an excuse to link it above, but maybe the draft page at CirrusSearch/Scoring that mentions some of the factors that go into retrieval and scoring, particularly for queries that can't be run through mjolnir models, might help as well.