Retrieving Information from Solr JOSA Data Science Bootcamp
● Head of Technology @ OpenSooq.com
● Technical Reviewer for “Scaling Apache Solr” and “Apache Solr Search Patterns” (Books)
● Contributor in Apache Solr● Built 10 search engines in the
last 2 years
Ramzi Alqrainy
Topics to be covered ● Exploring Solr’s Query Form● Basic Queries and Parameters● Matching Multiple Terms● Fuzzy Matching● Range Searches● Sorting● Pseudo Fields● Geospatial Searches● Filter Queries● Faceting and Stats● Tuning Relevance
Boolean Queries
● Search for two different terms, new and house,requiring both to match
● Search for two different terms, new and house, requiring only one to match
● Default operator is OR, can be changed using the q.op query parameter.
Inverted Index—Revisited
● All terms in the index map to 1 or more documents.● Terms in inverted index are stored in ascending
lexicographical order● When searching for multiple terms/ expressions, Solr (and
Lucene) returns multiple document result sets corresponding to the various terms in the query and then does the specified binary operations on these result sets in order to generate the final result set.
● Scoring is performed on the result set o generate final result
Proximity Searches● Represent arbitrarily complex queries● Solr/Lucene not only stores the documents that contain the
terms, but also their positions within a document (term positions), which is used to provide phrase and proximity search functionality
● The number of the “~” is called a slop factor and has a hard limit of 2, above which the number of permutations get too large to provide results within a reasonable time
Fuzzy Edit-Distance Searching
● Flexibility to handle misspellings and different spellings of a word
● Character variations based on Damerau-Levenshtein distances
● Accounts for 80% of human misspellings
Wildcard Matching
● Robust functionality, but can be expensive if not properly used.
○ First all terms that match parts of the term before wildcard expression are extracted
○ Then all those terms are inspected to see if they match the entire wildcard expression
○ Expensive if your expression matches a large number of terms (for example the query e*)
Query on a Range
● Solr Date Time uses a format that is a restricted form of the canonical representation of dateTime in the XML Schema specification (inspired by ISO 8601). All times are assumed to be UTC (no timezone specification)
● Based on a lexicographically sorted order for the field being queried● Solr has Trie field types (tint, tdate, etc.) that should be used when you are
doing a large number of range queries● Various field types will be covered later in the course
Solr Date Syntax
● Uses UTC and Restricted DateTime format● Allows rounding down by YEAR, MONTH, WEEK, DAY,
MINUTE, SECOND● NOW represents current time and using DateMath, we can
specify yesterday, tomorrow, last year, etc.
Pseudo Fields
● Dynamically added at query time and calculated from fields in the schema using in- built functions
● Through functions, you can manipulate the values of any field before it is returned
● Can also be used to modify the order of documents by sorting on the pseudo field
Geospatial Searches● Solr provides location-based search● Define a “location” field that contains latitude and longitude● You can use a Query parser called “geofilt” to search on this
field, specifying the point and radius around it● Another query parser bbox uses a square around the point to
do faster but approximate calculations● Other types of searches (grids, polygons, etc. are possible
and covered in advanced course
Returning Calculated Distances
● You can use a pseudo field (a field that is calculated at query time) to achieve this
The fq and q Parameters● Indistinguishable at first glance: same query parameters passed to either
parameters will return same documents.● But,
○ fq serves a single purpose, to limit what is returned○ q limits what is returned AND supplies the relevancy algorithm with a set
of terms used for scoring● fq results are cached and can be reused between searches● Using fq we can avoid unnecessary relevancy calculations● You can use multiple fq’s in a request (each individually cached), but only one q
parameter
Faceted Search● High-level breakdown of search
results based on one or more aspects (facets) of their documents
● Allows users to filter by (drill down into) specific components
● Can facet on values of fields, or facet by queries
Field Faceting
● Request back the unique values found in a particular field
● Most commonly used● Works for single- and multi-valued
fields● Values are based on the indexed
values of the field● Common practice is to facet on a
String field and search on a text field (to be discussed later). So, some schema preparation is required for faceting
Date Range Facets
● Recall Solr Date Syntax covered earlier in class● Uses UTC and Restricted DateTime format● Allows rounding down by YEAR, MONTH, WEEK, DAY,
MINUTE, SECOND● NOW represents current time and using DateMath, we can
specify yesterday, tomorrow, last year, etc.
Stats and Facets
● Can get aggregations on various fields● From Solr 5.x onwards, stats on pivot facets is also available● See https://lucidworks.com/blog/you-got-stats-in-my-facets/ for
a great explanation of faceting
Pivot Facets
● Functions like pivot tables in spreadsheet apps● Aggregate calculations that pivot on values from multiple fields● Example: give me a count of 3,4 and 5 star hotels in the top
three cities● Solr 5.x also allows you to stats calculations on pivots
Facet by Query
● Sometimes, you need unequal ranges ● You can use the facet.query parameter ● Provides counts for subqueries
Precision and recall
Precision and recall
Are the top results we show to users relevant?
Recall
Of the full set of documents found, have we found all of the relevant content in the index?
RelevancyOur goal is to give users relevant results Relevance is a soft or fuzzy thing
● Depends upon the judgment of users
Scoring is our attempt to predict relevance
Similarity classes hold the implementations
● DefaultSimilarity ( TF-IDF )● BM25Similarity● DFRSimilarity● IBSimilarity● LMDirichletSimilarity● LMJelinekMercerSimilarity
Lucene Scoring
Similarity scoring formula
• Used to rank results by measuring the similarity between a query and the documents that match the query
Domain knowledgeExamples
● Cheaper● Newer or more recent● More popular or higher user clicks Higher average user ratings
Interesting combinations
● Value = average user ratings ÷ price ● Staying power = recent popularity ÷ age
Boosting and biasing
Lucene uses a standardized scoring approach
Lucene does not know:
● Your data● Your users● Their queries Their preferences
Domain knowledge
What do you know about your data?
● Any specific rules about your data that wouldn't be suitable in a generic IR scoring algorithm
● In many data domains, there are fundamental numeric properties that make some objects generally "better" than others
Domain knowledgeMore subtle examples
● Novelty factor○ Quantity of user ratings × stdDev of ratings Profit margin
● Profit margin○ Retail price ‒ factory cost Scarcity
● Scarcity○ Quantity remaining
● Popularity by association or categorization○ Sweaters sell better then swimsuits in November
● Manual ranking○ New York Times bestseller list
How can you improve search results?
Using a sledge hammer
● Ignore score, sort on X
● Filter by X, retry if 0 results
How can you improve search results?
● Boost functions and queries● Apply domain knowledge based on numeric properties by
multiplying functions directly into the score