In this post, we will examine what sets Apache Solr aside as a search engine, from conventional databases like MongoDB. We will explore the similarities and differences between Solr and MongoDB by analyzing a series of comparative queries. We will then delve into some of Solr’s more advanced search capabilities.

Why Search?

The ability to search for information is a basic requirement of many applications. Architects and Developers who limit themselves to traditional databases, often attempt to meet search requirements by creating unnecessarily and overly complex SQL query-based solutions. They force end-users to search in unnatural or highly-structured ways or provide results that lack a sense of relevancy. End-users are not Database Administrators, they do not understand the nuances of SQL, they simply want relevant responses to their inquiries.

In a scenario where data consumers are arbitrarily searching for relevant information within a distinct domain, implementing a search-optimized, Lucene-based platform, such as Elasticsearch or Apache Solr, for reads, is often an effective solution.

Separating database reads from writes is not uncommon. I’ve worked on many projects where the requirements suggested an architecture in which one type of data storage technology should be implemented to optimize for writes, while a different type or types of data storage technologies should be implemented to optimize for reads. Architectures in which this is common include the following.

• CQRS (Command Query Responsibility Segregation) and Event Sourcing;
• Reporting, Data Analytics, and Big Data;
• ML (Machine Learning) and AI (Artificial Intelligence);
• Real-Time and Streaming Data (such as IoT);
• Search: Content, Document, Knowledge;

In this post, we will examine the search capabilities of Apache Solr. We will compare and contrast Solr’s search capabilities to those of MongoDB, the leading NoSQL database. We will consider the differences between querying for data and searching for information.

Apache Lucene

According to Apache, the Apache Lucene project develops open-source search software, including the following sub-projects: Lucene Core, Solr, and PyLucene. The Lucene Core sub-project provides Java-based indexing and search technology, as well as spellchecking, hit highlighting, and advanced analysis/tokenization capabilities.

Apache Lucene 7.7.0 and Apache Solr 7.7.0 were just released in February 2019 and used for all the post’s examples.

Apache Solr

According to Apache, Apache Solr is the popular, blazing fast, open source, enterprise search platform built on Apache Lucene. Solr powers the search and navigation features of many of the world’s largest internet sites.

Apache Solr includes the ability to set up a cluster of Solr servers that combines fault tolerance and high availability. Referred to as SolrCloud, and backed by Apache Zookeeper, these capabilities provide distributed, sharded, and replicated indexing and search capabilities.

According to Wikipedia, Solr was created at CNET Networks in 2004, donated to the Apache Software Foundation in 2006, and graduated from their incubator in 2007. Solr version 1.3 was released in 2008. In 2010, the Lucene and Solr projects merged; Solr became a Lucene subproject. With Apache Solr 7.7.0 just released, Solr has well over ten years of development and enterprise adoption behind it.

MongoDB

The leading NoSQL database, MongoDB, describes itself as a document database with the scalability and flexibility that you want with the querying and indexing that users need. Mongo features include ad hoc queries, indexing, and real-time aggregation, which provide powerful ways to access and analyze your data.

Released less than a year ago, MongoDB 4.0 added multi-document ACID transactions, data type conversions, non-blocking secondary replica reads, SHA-2 authentication, MongoDB Compass aggregation pipeline builder, Kubernetes integration, and the MongoDB Stitch serverless platform. MongoDB 4.0.6 was just released in February 2019 and used for all the post’s examples.

Comparing Search Features

Solr and MongoDB appear to have many search-related features in common.

• Both Solr and MongoDB are document-based data stores;
• Both Solr and MongoDB use a non-relational data model;
• Both feature advanced querying and indexing capabilities;
• Solr implements Lucene-based search capabilities; MongoDB has text-based search capabilities;
• Solr scores the relevance of search results using the Lucene scoring algorithm; MongoDB has the capability of ranking text search results using the $meta operator;
• Solr is able to selectively boost the relative importance of search fields and specific values in a field when calculating scores; MongoDB has the capability of boosting the relative importance of fields used in a text search using text indexes;
• Both Solr and MongoDB are capable of implementing stop words, stemming, and tokenization;

Demonstration

Source Code Examples

All examples shown in this post are available as a series of Python 3 scripts, contained in an open-source project on GitHub, searching-solr-vs-mongodb. The project contains the script, query_mongo.py, which uses the Python driver for MongoDB, pymongo, to execute all the MongoDB queries in this post. The project also contains the script, query_solr.py, which uses the lightweight Python wrapper for Apache Solr, pysolr, to execute all the Solr searches in this post. Both packages, along with ancillary packages, may be installed with pip.

pip3 install pysolr pymongo bson.json_util requests

MongoDB and Solr Instances

To follow along, you will need your own MongoDB and Solr instances. Both are easily stood up locally with Docker, using the official MongoDB and Solr Docker Hub images. Example docker run commands are shown below.

The second command, the Solr command, also creates a new Solr core. The command also bind-mounts the ‘conf’ directory, within the local project, into the container. This will give us the ability to modify our index’s configuration and to store that configuration in source control. All data is ephemeral, neither container persists data outside the container, using these particular commands.

docker run --name mongo -p 27017:27017 -d mongo:latest

docker run --name solr -d \
  -p 8983:8983 \
  -v $PWD/conf:/conf \
  solr:latest \
  solr-create -c movies -d /conf

Environment Variables

The source code expects two environment variables, which contain the connection information for MongoDB and Solr. You will need to replace the values below with your own connection strings if they are different than the examples below, used for Docker.

export SOLR_URL="http://localhost:8983/solr"
export MONOGDB_CONN="mongodb://localhost:27017/movies"

Importing Movies to MongoDB

For this post, we will be using a publicly available movie dataset from MongoDB. A copy of the dataset is available in the project, as well as on MongoDB’s website, Setup and Import the Data.

Assuming you have an instance of MongoDB accessible and have set the two environment variables above, import the specially-formatted JSON file for MongoDB, movieDetails_mongo.json, directly into the movies database’s movieDetails collection, using the following mongoimport command.

mongoimport \
  --uri $MONOGDB_CONN \
  --collection "movieDetails" \
  --drop --file "data/movieDetails_mongo.json"

Below is a view of the movies database’s movieDetails collection, running in the Docker container, as shown in the MongoDB Compass application.

Indexing Movies to Solr

Assuming you have an instance of Solr accessible and have set the two environment variables above, import the contents of the JSON file, movieDetails.json, by running the Python script, solr_index_movies.py, using the following command.

python3 ./solr_index_movies.py

The command executes a series of HTTP calls to Solr’s exposed RESTful API.

Below is a view of the Solr Administration User Interface, running within the Docker container, and showing the new movies core. After running the script, we should have 2,250 movie documents indexed.

The Solr Admin UI offers a number of useful tools for examing indexes, reviewing schemas and field types, and creating, analyzing, and debugging Solr queries. Below we see the Query UI with the results of a query displayed.

Tuning the Solr Index

The movies index uses a default schema, which was created when the movie documents were indexed. To optimize our query results, we will want to make a few adjustments to the default movies schema. First, we want to ensure that our Solr searches consider the pluralization of words. For example, when we search for the search term ‘Adventure’, we want Solr to also return documents containing terms like adventure, adventures, adventure’s, adventuring, and adventurer, but not misadventure. This is known as Stemming, or reducing words to their word stem. An example is shown below in the Solr Analysis UI.

The fields that we want to search, such as title, plot, and genres, were all indexed by default as the ‘text_general’ Solr field type. The ‘text_general’ field type does not implement stemming when indexing or querying. We need to switch the title, plot, and genres fields to the ‘text_en’ (English text) field type. The ‘text_en’ field type implements multiple indexing and querying filters, including the PorterStemFilterFactory filter, which removes common endings from words. Similar filters include the English Minimal Stem Filter and the English Possessive Filter.

Additionally, the MultiValued field property is set to true by default for these fields in Solr. Since the title and plot fields, amongst others, were only intended to hold a single text value, as opposed to an array of values, we will switch the MultiValued field property to false. This helps with sorting and filtering, and the correct deserialization of documents.

The solr_index_movies.py script will change the title, plot, and genres fields from text_general’ to ‘text_en’ and change the title and plot fields from multi-valued to single-valued. Since we have changed the index’s schema, the script will re-import all the documents after making the schema changes.

To get a better sense of what the schema changes look like, let’s look at the equivalent cURL command to change the schema. This gives you a better sense of the field-level modifications we are making.

You can use the Schema UI to view the results, as shown below. Note the new field types for title, plot, and genres. Also, note the index and query analyzers, including the  PorterStemFilterFactory, used by the ‘text_en’ field type.

Comparative Queries

To demonstrate the similarities and the differences between Solr and MongoDB, we will examine a series of comparative queries, followed by a series of Solr-only searches. Again, all queries and output shown are included in the two project’s Python scripts.

Query 1a: All Documents

To start, we will perform a simple query for all the movie documents in the MongoDB collection, followed by the Solr index. With MongoDB, we use the find method. With Solr, we will use the Standard Query Parser, commonly known as the ‘lucene’ query parser, and the q (query) parameter. The result of the queries should be identical, with all 2,250 documents returned.

MongoDB:

Parameters
----------
query: {}
  
Results
----------
document count: 2250

Solr:

Parameters
----------
q: *:*
kwargs: {}
  
Results
----------
document count: 2250

Query 1b: Count Only

We can alter our first query to limit our response to only the count of documents for a given query in MongoDB; no documents will be returned. Since our query is empty, we will get back a count of all documents in the MongoDB database’s collection.

db.movieDetails.count()

Similarly, in Solr, we can set the rows parameter to zero to return only the document count. For brevity, we can also omit the Solr response header using the omitHeader  parameter.

Parameters
----------
q: *:*
kwargs: {
  'omitHeader': 'true', 
  'rows': '0'
}

Results
----------
document count: 2250

Query 2: Exact Search

Next, we will perform a query for the exact movie title, ‘Star Wars: Episode V – The Empire Strikes Back’ in MongoDB, then Solr. Again, the results of the queries should be identical, with one document returned, matching the title.

MongoDB:

Parameters
----------
query: {'title': 'Star Wars: Episode V - The Empire Strikes Back'}
projection: {'_id': 0, 'title': 1}
  
Results
----------
document count: 1
{'title': 'Star Wars: Episode V - The Empire Strikes Back'}

The quotes around the title are key for Solr to view the query as a single phrase as opposed to a series of search terms.

Solr:

Parameters
----------
q: title:"Star Wars: Episode V - The Empire Strikes Back"
kwargs: {
  'defType': 'lucene', 
  'fl': 'title, score'
}
  
Results
----------
document count: 1
{'title': 'Star Wars: Episode V - The Empire Strikes Back', 'score': 29.41}

Note the use of 'defType': 'lucene' is optional. The standard Lucene query parser is the default parser used by Solr. I am merely showing this parameter to improve the reader’s understanding. Later, we will use other query parsers.

Query 3: Search Phrase

Next, we will perform a query for the phrase ‘star wars’. With MongoDB, we will use the $regex and $options Evaluation Query Operators. The results from both the MongoDB and Solr queries should be identical, the six Star Wars movies are returned.

MongoDB:

Parameters
----------
query: {
  'title': {'$regex': '\\bstar wars\\b', 
  '$options': 'i'}
}
projection: {'_id': 0, 'title': 1}
  
Results
----------
document count: 6
{'title': 'Star Wars: Episode I - The Phantom Menace'}
{'title': 'Star Wars: Episode II - Attack of the Clones'}
{'title': 'Star Wars: Episode III - Revenge of the Sith'}
{'title': 'Star Wars: Episode IV - A New Hope'}
{'title': 'Star Wars: Episode V - The Empire Strikes Back'}
{'title': 'Star Wars: Episode VI - Return of the Jedi'}

With Solr, wrapping the phrase ‘star wars’ in quotes ensures Solr will treat the query string as an exact phrase, not individual search terms. Solr results are scored, but scores are almost all identical since all six movies contain the exact phrase.

Solr:

Parameters
----------
q: "star wars"
kwargs: {
  'defType': 'lucene', 
  'df': 'title', 
  'fl': 'title, score'
}
  
Results
----------
document count: 6
{'title': 'Star Wars: Episode VI - Return of the Jedi', 'score': 8.21}
{'title': 'Star Wars: Episode II - Attack of the Clones', 'score': 8.21}
{'title': 'Star Wars: Episode IV - A New Hope', 'score': 8.21}
{'title': 'Star Wars: Episode I - The Phantom Menace', 'score': 8.21}
{'title': 'Star Wars: Episode III - Revenge of the Sith', 'score': 8.21}
{'title': 'Star Wars: Episode V - The Empire Strikes Back', 'score': 7.55}

Here is the actual Lucene query (q) Solr will run.

title:"star war"

Query 4: Search Terms

Next, we will perform a query for all movies whose title contains either the search terms ‘star’ or ‘wars’, as opposed to the phrase, ‘star wars’. The Solr web console has a very powerful Analysis tool. Using the Analysis tool, we can examine how each filter (abbreviated in the far left column, below), associated with a particular field type, will impact the matching capabilities of Solr. To use the Analysis tool, place your search term(s) or phrase on the right side, an indexed field value on the left, and choose a field or field type from the dropdown.

Below, we see how the search terms ‘Star’ and ‘Wars’ (shown below right) would match a series of variations on the two words (shown below left) if the fields being searched are of the field type, ‘text_en’, as discussed earlier. For example, a query for ‘Star’ would match ‘star’, ‘stars’, ‘star‘s’, ‘starring’, ‘starred’, ‘star-shaped’, but not ‘shaped’, ‘superstars’, ‘started’.

Below, we see similar results for the search term, ‘Wars’.

MongoDB Text Search

To accomplish the query with MongoDB, we will use MongoDB’s $text Evaluation Query Operator. The MongoDB $text operator is able to perform a text search across multiple fields indexed with a text index. MongoDB’s text indexes support text search queries on string content. The text indexes can include any field whose value is a string or an array of string elements, such as the movieDetail collection’s genres field. Although not as powerful as Solr’s search capabilities, MongoDB’s text search may address many basic search requirements without the need to augment the architecture with a search engine.

For our next query, we will rely on a text index on the title field. When the Python script runs, it creates the following three indexes on the collection, including the title text index.

With the text index in place, the result of the queries should be identical, with 18 documents returned. Both the MongoDB and Solr resultsets are scored, however, both are scored differently, using different algorithms.

MongoDB:

Parameters
----------
query: {
  '$text': {'$search': 'star wars', '$language': 'en', '$caseSensitive': False}, 
  'countries': 'USA'
}
projection: {'score': {'$meta': 'textScore'}, '_id': 0, 'title': 1}
sort: [('score', {'$meta': 'textScore'})]
  
Results
----------
document count: 18
{'title': 'Star Wars: Episode I - The Phantom Menace', 'score': 1.2}
{'title': 'Star Wars: Episode IV - A New Hope', 'score': 1.17}
{'title': 'Star Wars: Episode VI - Return of the Jedi', 'score': 1.17}
{'title': 'Star Wars: Episode II - Attack of the Clones', 'score': 1.17}
{'title': 'Star Wars: Episode III - Revenge of the Sith', 'score': 1.17}

Solr:

Parameters
----------
q: star wars
kwargs: {
  'defType': 'lucene', 
  'fq': 'countries: USA' 
  'df': 'title', 
  'fl': 'title, score', 
  'rows': '5'
}
  
Results
----------
document count: 18
{'title': 'Star Wars: Episode VI - Return of the Jedi', 'score': 8.21}
{'title': 'Star Wars: Episode II - Attack of the Clones', 'score': 8.21}
{'title': 'Star Wars: Episode IV - A New Hope', 'score': 8.21}
{'title': 'Star Wars: Episode I - The Phantom Menace', 'score': 8.21}
{'title': 'Star Wars: Episode III - Revenge of the Sith', 'score': 8.21}

Here is the actual Lucene query (q) Solr will run. The countries filter is applied afterward.

title:star title:war

Find me a Good Western

Query 5a: Multiple Search Terms

Next, we will perform a query for movies, produced in the USA, with the search terms ‘western’, ‘action’, or ‘adventure’ in the movie genres field. The genres field may hold multiple genre values. Although this is a simple query, we can start to see the advantages of Solr’s Lucene scoring capability to provide a way to measure the relevancy of individual results.

Even limited to the USA-based movies, this genres query returns a large number of results, 244 documents. With MongoDB, we have no sense of which documents are more relevant than others. Compared to the Solr results, MongoDB got a few in the top five results, but not the most relevant, based on matching all or most of the genres.

MongoDB:

Parameters
----------
query: {
  'genres': {'$in': ['Adventure', 'Action', 'Western']}, 
  'countries': 'USA'
}
projection: {'_id': 0, 'genres': 1, 'title': 1}
  
Results
----------
document count: 244
{'title': 'Wild Wild West', 'genres': ['Action', 'Western', 'Comedy']}
{'title': 'A Million Ways to Die in the West', 'genres': ['Comedy', 'Western']}
{'title': 'An American Tail: Fievel Goes West', 'genres': ['Animation', 'Adventure', 'Family']}
{'title': 'Once Upon a Time in the West', 'genres': ['Western']}
{'title': 'How the West Was Won', 'genres': ['Western']}

However, with Solr’s scoring, we see the first (top) result, ‘The Wild Bunch’, has a score of 7.18. It genres contain exactly ‘western’, ‘action’, or ‘adventure’. The last (bottom) result, ‘S.S. Doomtrooper’, has a score of 1.47. The most relevant result scored nearly 5x higher (488%) than the least most relevant result. If you were searching for a western action adventure movie, it is pretty apparent the top Solr result, ‘The Wild Bunch’, is a much better choice than the bottom result, ‘S.S. Doomtrooper’. In fact, as shown below, all five top-scoring Solr results look pretty promising based on their score, genres, and title.

Solr:

Parameters
----------
q: adventure action western
kwargs: {
  'defType': 'lucene', 
  'fq': 'countries: USA', 
  'df': 'genres', 
  'fl': 'title, genres, score', 
  'rows': '5'
}
  
Results
----------
document count: 244
{'title': 'The Wild Bunch', 'genres': ['Action', 'Adventure', 'Western'], 'score': 7.18}
{'title': 'Crossfire Trail', 'genres': ['Action', 'Western'], 'score': 6.26}
{'title': 'The Big Trail', 'genres': ['Adventure', 'Western', 'Romance'], 'score': 5.46}
{'title': 'Once Upon a Time in the West', 'genres': ['Western'], 'score': 5.26}
{'title': 'How the West Was Won', 'genres': ['Western'], 'score': 5.26}

Here is the actual Lucene query (q) Solr will run. The countries filter is applied afterward.

genres:adventur genres:action genres:western

Query 5b: Required Search Term

There are nearly endless options that can be used with Solr to influence Solr’s results. For example, we could perform the same Solr query above, but this time require that the word ‘western’ is the genres field, by using the plus symbol (+) boolean operator. The top five results and scores are the same, but the total number of relevant results have decreased from 244 to just 24. That means 220 of the previous results contained ‘action’, and/or ‘adventure’, but not ‘western’. The opposite is also true, using the minus symbol (-) boolean operator will ensure the results do not contain a particular word or phrase.

Solr:

Parameters
----------
q: adventure action +western
kwargs: { 
  'defType': 'lucene', 
  'fq': 'countries: USA', 
  'df': 'genres', 
  'fl': 'title, genres, score', 
  'rows': '5'
} 
  
Results
----------
document count: 24
{'title': 'The Wild Bunch', 'genres': ['Action', 'Adventure', 'Western'], 'score': 7.18}
{'title': 'Crossfire Trail', 'genres': ['Action', 'Western'], 'score': 6.26}
{'title': 'The Big Trail', 'genres': ['Adventure', 'Western', 'Romance'], 'score': 5.46}
{'title': 'Once Upon a Time in the West', 'genres': ['Western'], 'score': 5.26}
{'title': 'How the West Was Won', 'genres': ['Western'], 'score': 5.26}

Here is the actual Lucene query (q) Solr will run. The countries filter is applied afterward.

(genres:adventur genres:action) +genres:western

Query 6a: eDisMax Query

For our next query, we will compare Solr’s eDisMax query parser to MongoDB’s text search capabilities.

Solr Extended DisMax

According to Solr, The DisMax query parser is designed to process simple phrases and to search for individual terms across several fields using different weighting (boosts) based on the significance of each field. Additional options enable users to influence the score based on rules specific to each use case (independent of user input). Solr’s Extended DisMax (eDisMax) query parser is an improved version of the DisMax query parser.

In my opinion, in addition to the Lucene-based scoring, the ability to easily search across multiple fields and selectively boost results with the DisMax and eDisMax query parsers is what starts to differentiate querying data in a database, from searching for relevant results with a search engine.

Multi-Field Text Index

For our next query, the Python script will drop the previous MongoDB text index on the title field and create a new compound text index, which will incorporate the title, plot, and genres fields.

Below, we see the new compound text index in the MongoDB Compass application’s Indexes tab.

We will perform a query for movies, produced in the USA, with the search terms ‘western’, ‘action’, or ‘adventure’ in the movie title, plot, or genres fields. The results of the queries should be identical, with 259 documents returned. Both the MongoDB and Solr resultsets are scored, but again, the scores and ordering of results are not identical. Of the top ten results, the two queries matched six movies in their top ten results.

MongoDB:

Parameters
----------
query: {
  '$text': {'$search': 'western action adventure', '$language': 'en', '$caseSensitive': False}, 
  'countries': 'USA'
}
projection: {'score': {'$meta': 'textScore'}, '_id': 0, 'title': 1}
  
Results
----------
document count: 259
{'title': 'Zathura: A Space Adventure', 'genres': ['Action', 'Adventure', 'Comedy'], 'score': 3.3}
{'title': 'The Extraordinary Adventures of Adèle Blanc-Sec', 'genres': ['Action', 'Adventure', 'Fantasy'], 'score': 3.24}
{'title': 'The Wild Bunch', 'genres': ['Action', 'Adventure', 'Western'], 'score': 3.2}
{'title': 'The Adventures of Tintin', 'genres': ['Animation', 'Action', 'Adventure'], 'score': 2.85}
{'title': 'Adventures in Babysitting', 'genres': ['Action', 'Adventure', 'Comedy'], 'score': 2.85}

Solr:

Parameters
----------
q: western action adventure
kwargs: {
  'defType': 'edismax', 
  'fq': 'countries: USA', 
  'qf': 'plot title genres', 
  'fl': 'title, genres, score', 
  'rows': '5'
}
  
Results
----------
document count: 259
{'title': 'The Secret Life of Walter Mitty', 'genres': ['Adventure', 'Comedy', 'Drama'], 'score': 7.67}
{'title': 'Western Union', 'genres': ['History', 'Western'], 'score': 7.39}
{'title': 'The Adventures of Tintin', 'genres': ['Animation', 'Action', 'Adventure'], 'score': 7.36}
{'title': 'Adventures in Babysitting', 'genres': ['Action', 'Adventure', 'Comedy'], 'score': 7.36}
{'title': 'The Poseidon Adventure', 'genres': ['Action', 'Adventure', 'Drama'], 'score': 7.36}

Here is the actual Lucene query Solr will run.

+(
  (title:adventur | plot:adventur | genres:adventur) 
  (title:action | plot:action | genres:action) 
  (title:western | plot:western | genres:western)
)

Query 6b: Boosting Fields

If you really wanted a ‘western action adventure’ movie as opposed to a ‘western’, ‘action’, or ‘adventure’ movie, then neither Solr or MongoDB’s probably completely satisfied you with their first five search results. Boosting or weighting fields can often provide more relevant search results if the correct fields are boosted, and the amount of positive or negative boost is appropriate.

MongoDB’s text indexes also allow for weighting individual fields. The weight of an indexed field denotes the significance of the field relative to the other indexed fields and directly impacts the text search score. Weighting fields are the equivalent to boosting fields with Solr. Below, we see a modification applied to our previous text index in which the title field is given twice the weight of the plot field (1.0 is default) and the genres field is given twice the weight of the title field. The Python script also applies this index for you.

Likewise, Solr is also capable of boosting fields for both the DisMax and eDisMax query parsers. For our next query, we will repeat the previous query, but boost fields in the eDisMax’s qf (Query Fields) parameter to match the boost in the MongoDB weighted text index, shown above.

The results of the queries should be identical, with 259 documents returned. MongoDB and Solr’s results are scored and ordered differently. However, compared to the previous, un-weighted/boosted MongoDB and Solr query results above, the relative scores are higher, the order of movies returned are different, and most importantly, the Solr results seem more relevant to the original search intent.

For example with Solr, take the movie, ‘The Secret Life of Walter Mitty’, which previously scored highest at 7.58. In the boosted search results, the movie, ‘The Wild Bunch’ is now ranked first with a score of 28.71. The movie, ‘The Secret Life of Walter Mitty’ is no longer even in the top 50 Solr results. Comparatively, other movies, like ‘The Adventures of Tintin’ and ‘Adventures in Babysitting’, barely moved in position even though their scores changed proportionally.

MongoDB:

Parameters
----------
query: {
  '$text': {'$search': 'western action adventure', '$language': 'en', '$caseSensitive': False}, 
  'countries': 'USA'
}
projection: {'score': {'$meta': 'textScore'}, '_id': 0, 'title': 1}
  
Results
----------
document count: 259
{'title': 'The Wild Bunch', 'genres': ['Action', 'Adventure', 'Western'], 'score': 12.8}
{'title': 'Zathura: A Space Adventure', 'genres': ['Action', 'Adventure', 'Comedy'], 'score': 10.27}
{'title': 'The Extraordinary Adventures of Adèle Blanc-Sec', 'genres': ['Action', 'Adventure', 'Fantasy'], 'score': 10.14}
{'title': 'The Adventures of Tintin', 'genres': ['Animation', 'Action', 'Adventure'], 'score': 9.9}
{'title': 'Adventures in Babysitting', 'genres': ['Action', 'Adventure', 'Comedy'], 'score': 9.9}

Solr:

Parameters
----------
q: western action adventure
kwargs: {
  'defType': 'edismax', 
  'fq': 'countries: USA', 
  'qf': 'plot title^2.0 genres^4.0', 
  'fl': 'title, genres, score', 
  'rows': '5'
}
  
Results
----------
document count: 259
{'title': 'The Wild Bunch', 'genres': ['Action', 'Adventure', 'Western'], 'score': 28.71}
{'title': 'Crossfire Trail', 'genres': ['Action', 'Western'], 'score': 25.05}
{'title': 'The Big Trail', 'genres': ['Adventure', 'Western', 'Romance'], 'score': 21.84}
{'title': 'Once Upon a Time in the West', 'genres': ['Western'], 'score': 21.05}
{'title': 'How the West Was Won', 'genres': ['Western'], 'score': 21.05}

Here is the actual Lucene query Solr will run.

+(
  ((title:adventur)^2.0 | plot:adventur | (genres:adventur)^4.0) 
  ((title:action)^2.0 | plot:action | (genres:action)^4.0) 
  ((title:western)^2.0 | plot:western | (genres:western)^4.0)
)

Query 6c: eDisMax Boosted with Required/Prohibited Terms

We can use both the plus (+) and minus (-) boolean operators to obtain more relevant search results. Let’s repeat the last Solr boosted query, but this time, also require any results to contain the search term, ‘western’, and prohibit the responses from containing the search term, ‘romance’. I would consider the new search results, based these modifications to the Solr query, to be more relevant to the original intent of the search, than the previous results. For example, the movie, ‘The Big Trail’, a romantic western adventure movie, according to its genres, is no longer included in the results sets.

Solr:

Parameters
----------
q: adventure action +western -romance
kwargs: {
  'defType': 'edismax', 
  'fq': 'countries: USA', 
  'qf': 'plot title^2.0 genres^4.0',
  'fl': 'title, genres, score', 
  'rows': '5'
}

Results
----------
document count: 25
{'title': 'The Wild Bunch', 'genres': ['Action', 'Adventure', 'Western'], 'score': 28.71}
{'title': 'Crossfire Trail', 'genres': ['Action', 'Western'], 'score': 25.05}
{'title': 'Once Upon a Time in the West', 'genres': ['Western'], 'score': 21.05}
{'title': 'How the West Was Won', 'genres': ['Western'], 'score': 21.05}
{'title': 'Cowboy', 'genres': ['Western'], 'score': 21.05}

Here is the actual Lucene query Solr will run.

+(
  ((title:adventur)^2.0 | plot:adventur | (genres:adventur)^4.0) 
  ((title:action)^2.0 | plot:action | (genres:action)^4.0) 
  +((title:western)^2.0 | plot:western | (genres:western)^4.0) 
  -((title:romanc)^2.0 | plot:romanc | (genres:romanc)^4.0)
)

Query 7a: The Movie Dilemma

Frequently, end-users interact with a search engine, such as Google, through a search box. We type something into a search box and get back a list of relevant results. By now, most of us have learned how to phrase our Google search to get optimal results. However, the reality is, people can and will type just about anything into a search box.

To try and improve the average search results for end-users, search engineers will often try to tune query parameters, such as boosting the importance certain search fields over other, adjusting fuzzy search parameters, or ignore irrelevant words in the search phases by adding them to the stop words. Default English in Solr stop words include like: ‘a’, ‘an’, ‘and’, ‘are’, ‘as’, ‘at’, ‘be’, ‘but’, ‘by’, ‘for’, and so forth.

Take, for example, the word ‘movie’. Someone searching for a movie to watch, using a search box, might enter the phrase ‘A cowboy movie’. The term, ‘A’, is ignored as a stop word. This leaves the search terms ‘cowboy’ and ‘movie’ to be searched for in the title, plot, and genres fields. As we see by the top ten results below, most appear to be about cowboys, having the word ‘cowboy’ in their title or plot. Then there is the result, ‘TV: The Movie’. This does not appear to be a movie about cowboys. The word ‘cowboy’ is not in the title, plot, or genres, yet here it is in third place since it contains the word ‘movie’ in the title, plot, and/or genres.

Similarly, the top movie result, ‘Cowboy Bebop: The Movie’, is probably no more relevant than the second, third, or fourth place movies. However, ‘Cowboy Bebop: The Movie’ has scored significantly higher than even the number two results (11.24 vs. 7.31). This is because the movie’s title contains both search terms, ‘cowboy’ and ‘movie’, even though the word ‘movie’ is irrelevant to the original intent of the search phrase.

Solr:

Parameters
----------
q: A cowboys movie
kwargs: {
  'defType': 'edismax', 
  'fq': 'countries: USA', 
  'qf': 'plot title genres', 
  'fl': 'title, genres, score', 
  'rows': '10'
}
  
Results
----------
document count: 23
{'title': 'Cowboy Bebop: The Movie', 'genres': ['Animation', 'Action', 'Crime'], 'score': 11.24}
{'title': 'Cowboy', 'genres': ['Western'], 'score': 7.31}
{'title': 'TV: The Movie', 'genres': ['Comedy'], 'score': 6.42}
{'title': 'Space Cowboys', 'genres': ['Action', 'Adventure', 'Thriller'], 'score': 6.33}
{'title': 'Midnight Cowboy', 'genres': ['Drama'], 'score': 6.33}
{'title': 'Drugstore Cowboy', 'genres': ['Crime', 'Drama'], 'score': 6.33}
{'title': 'Urban Cowboy', 'genres': ['Drama', 'Romance', 'Western'], 'score': 6.33}
{'title': 'The Cowboy Way', 'genres': ['Action', 'Comedy', 'Drama'], 'score': 6.33}
{'title': 'The Cowboy and the Lady', 'genres': ['Comedy', 'Drama', 'Romance'], 'score': 6.33}
{'title': 'Toy Story', 'genres': ['Animation', 'Adventure', 'Comedy'], 'score': 5.65}

Here is the actual Lucene query Solr will run.

+(
  (title:cowboi | plot:cowboi | genres:cowboi) 
  (title:movi | plot:movi | genres:movi)
)

Query 7b: Stop Words

To solve the movie dilemma, we might consider adding the word ‘movie’ to the stop words, since the word ‘movie’ seems to be irrelevant to the search phrase ‘A cowboys movie’, or to a movie search engine in general. However, this choice will negatively impact other searches. There are 12 movie titles containing the word ‘movie’. If you were searching for ‘The Lego Movie’, ignoring the word ‘movie’, as a stop word, would negatively impact the accuracy and relevance of your search results. You would end up with only one of two Lego movies in your search results, the one without the title that contained the word ‘movie’. Note only one of the two Lego movies is returned.

Solr:

Parameters
----------
q: The Lego Movie
kwargs: {
  'defType': 'edismax', 
  'fq': 'countries: USA', 
  'qf': 'plot title genres', 
  'fl': 'title, genres, score', 
  'rows': '5'
}
  
Results
----------
document count: 1
{'title': 'Lego DC Comics Super Heroes: Justice League vs. Bizarro League', 'genres': ['Animation', 'Action', 'Adventure'], 'score': 3.63}

Here is the actual Lucene query Solr will run. Note neither the term ‘a’ and ‘movie’ are part of the search.

+((title:lego | plot:lego | genres:lego)

Query 7c: Negative Boost

A second method to solve the movie dilemma might be to negatively boost the word ‘movie’ when it appears in the title field, thus reducing its relevance. Negatively boosting fields, or more precisely, negatively boosting a specific field value, is possible with both the DisMax and eDisMax query parsers. We can assign a negative boost to the word ‘movie’ when it appears in the title field, by using the bq (Boost Query) parameter.  According to Solr, The bq parameter specifies an additional, optional, query clause that will be added to the user’s main query to influence the score. As Developers, we could programmatically append the negatively boosted term(s) into the query without directly altering the user’s original search phrase. Again, like stop words, boosting may also negatively impact other searches.

After some experimentation, we will try a boost value of -2. We still get 23 results, however, the top ten results now appear to be more relevant, based on the intent of our search phrase. The movies, ‘Cowboy Bebop: The Movie’ and ‘TV: The Movie’ are not present in the top ten results; their scoring was lowered. You can repeat this process for more search terms, like ‘movie’, positively or negatively boosting their scores, to improve the relevancy of the results.

Solr:

Parameters
----------
q: A cowboys movie
kwargs: {
  'defType': 'edismax', 
  'fq': 'countries: USA', 
  'qf': 'plot title genres', 
  'bq': 'title:movie^-2.0', 
  'fl': 'title, genres, score', 
  'rows': '10'
}
  
Results
----------
document count: 23
{'title': 'Cowboy', 'genres': ['Western'], 'score': 7.31}
{'title': 'Space Cowboys', 'genres': ['Action', 'Adventure', 'Thriller'], 'score': 6.33}
{'title': 'Midnight Cowboy', 'genres': ['Drama'], 'score': 6.33}
{'title': 'Drugstore Cowboy', 'genres': ['Crime', 'Drama'], 'score': 6.33}
{'title': 'Urban Cowboy', 'genres': ['Drama', 'Romance', 'Western'], 'score': 6.33}
{'title': 'The Cowboy Way', 'genres': ['Action', 'Comedy', 'Drama'], 'score': 6.33}
{'title': 'The Cowboy and the Lady', 'genres': ['Comedy', 'Drama', 'Romance'], 'score': 6.33}
{'title': 'Toy Story', 'genres': ['Animation', 'Adventure', 'Comedy'], 'score': 5.65}
{'title': "Ride 'Em Cowboy", 'genres': ['Comedy', 'Western', 'Musical'], 'score': 5.58}
{'title': "G.M. Whiting's Enemy", 'genres': ['Mystery'], 'score': 5.32}

Here is the actual Lucene query Solr will run, accounting for the negative boost.

+((title:cowboi | plot:cowboi | genres:cowboi) 
  (title:movi | plot:movi | genres:movi)) 
(title:movi)^-2.0

Function Query

Solr’s Lucene scoring is effective, but what if we wanted to use an additional, subjective measure of relevance to enrich our search results? If we examine the movies index schema, we will see there are quite a few rating-related fields that provide a sense of the movie’s quality, as judged by viewers and organizations. For example, there is a nested ‘tomato’ object, containing a number of qualitative data fields, such as the tomato rating and a tomato user rating. Tomato refers to Rotten Tomatoes. According to their site, Rotten Tomatoes provides the world’s most trusted recommendation resources for quality entertainment. Additionally, the movies index schema includes similar ‘awards’ and ‘imdb’ objects and a ‘metacritic’ rating. Here is a snippet of data from the movie, ‘Butch Cassidy and the Sundance Kid’, showing many of those qualitative fields.

  "imdb": {
    "id": "tt0064115",
    "rating": 8.1,
    "votes": 142642
  },
  "tomato": {
    "meter": 89,
    "image": "certified",
    "rating": 8.2,
    "reviews": 46,
    "fresh": 41,
    "consensus": "With its iconic pairing of Paul Newman and Robert Redford, jaunty screenplay and Burt Bacharach score, Butch Cassidy and the Sundance Kid has gone down as among the defining moments in late-'60s American cinema.",
    "userMeter": 93,
    "userRating": 4,
    "userReviews": 70088
  },
  "metacritic": 58,
  "awards": {
    "wins": 16,
    "nominations": 14,
    "text": "Won 4 Oscars. Another 16 wins & 14 nominations."
  }

According to Apache, Lucene scoring is a combination of the Vector Space Model (VSM) of Information Retrieval and the Boolean model. Lucene allows influencing search results by ‘boosting’. Using the Solr’s Function Query, we can apply a document-level multiplicative boost function, which will alter the scores of the query’s search results.

Query 8: Boost Function

If you remember in our previous example, we queried for ‘adventure action +western -romance’. If we run it again, without the boosted fields, we got back 25 documents, of which ‘Western Union’ ranked highest, with a score of 7.39.

Solr:

Parameters
----------
q: adventure action +western -romance
kwargs: {
  'defType': 'edismax', 
  'fq': 'countries: USA', 
  'qf': 'plot title genres', 
  'fl': 'title, awards.wins, score', 
  'rows': '5'
}

Results
----------
document count: 25
{'title': 'Western Union', 'awards.wins': [0.0], 'score': 7.39}
{'title': 'The Wild Bunch', 'awards.wins': [5.0], 'score': 7.18}
{'title': 'Western Spaghetti', 'awards.wins': [2.0], 'score': 6.64}
{'title': 'Crossfire Trail', 'awards.wins': [1.0], 'score': 6.26}
{'title': 'Butch Cassidy and the Sundance Kid', 'awards.wins': [16.0], 'score': 6.23}

Here is the actual Lucene query Solr will run.

+(
  (title:adventur | plot:adventur | genres:adventur) 
  (title:action | plot:action | genres:action) 
  +(title:western | plot:western | genres:western) 
  -(title:romanc | plot:romanc | genres:romanc)
)

Now, we will apply a boost function. In the function below, I have arbitrarily taken the number of awards won by each movie and divided it in half. The function is applied to the eDisMax’s boost parameter.

div(field(awards.wins,min),2)

This function has a multiplicative effect on the Lucene scoring of the documents in the result set, by boosting scores in proportion to the number of awards each movie has won. We now get movies that are a blend of both relevant results, based on our search phrase, as well as those movies that are highly acclaimed.

The impact of the multiplicative boost function is immediately apparent with the top result, ‘Butch Cassidy and the Sundance Kid’. This widely acclaimed movie climbed from fifth place in the previous search to first place, using the boost formula. The movie, ‘Butch Cassidy and the Sundance Kid’, won an amazing 16 awards, which is 16 more than that of the previous first place movie, ‘Western Union’, which won no awards, moving it down all the way down to 17th place in the boosted results. A movie that wasn’t even in the top five results, ‘Wild Wild West’, is now in second place, having received ten awards.

The impact of the boost function is most apparent in the scores. Previously, the score delta between the first and fifth positions in the results was 1.16. The delta between the first and second position was a mear 0.21. Now, with the boost function applied, the range of scoring, and consequently, the two deltas increased significantly, 36.78 compared to 1.16 and 23.83 compared to 0.21.

Solr:

Parameters
----------
q: adventure action +western -romance
kwargs: {
  'defType': 'edismax', 
  'fq': 'countries: USA', 
  'qf': 'plot title genres', 
  'fl': 'title, awards.wins, score', 
  'boost': 'div(field(awards.wins,min),2)', 
  'rows': '5'
}

Results
----------
document count: 25
{'title': 'Butch Cassidy and the Sundance Kid', 'awards.wins': [16.0], 'score': 49.86}
{'title': 'Wild Wild West', 'awards.wins': [10.0], 'score': 26.03}
{'title': 'How the West Was Won', 'awards.wins': [7.0], 'score': 18.42}
{'title': 'The Wild Bunch', 'awards.wins': [5.0], 'score': 17.95}
{'title': 'All Quiet on the Western Front', 'awards.wins': [5.0], 'score': 13.08}

Here is the actual Lucene query Solr will run, using the boost function.

FunctionScoreQuery(+((title:adventur | plot:adventur | 
  genres:adventur) (title:action | plot:action | genres:action) 
  +(title:western | plot:western | genres:western) 
  -(title:romanc | plot:romanc | genres:romanc)), 
  scored by boost(div(double(awards.wins,MIN),const(2))))

Solr’s Function Query offers a large number of mathematical functions, which can be combined into complex formulas. For example, we could also take the square root of the sum of the IMDB rating and the number of award nominations.

sqrt(sum(field(imdb.rating,min),field(awards.wins,min)))

More Like This, Please

You’ve found a good movie, and now you want more movies just like that one. Maybe you want more movies by a particular director, or starring a certain actor, or based on the same theme. Solr has a solution for this, the More Like This Query Parser. The MLTQParser, for short, enables retrieving documents that are similar to a given document. It uses Lucene’s existing MoreLikeThis logic. To use the parser, you provide the unique Solr ID of the document you want to find more like and the field(s) to use for the comparison. For example:

q: {!mlt qf="genres"}da54520e-a013-4ea3-9698-230ed02c8cf0

Query 9a: MLT Genres

In the first MLTQParser query, we will select the movie, ‘Star Wars: Episode I – The Phantom Menace’. We will look for more movies, produced in the USA, that are similar to ‘Star Wars: Episode I – The Phantom Menace’, by looking for similarities in the genres field. The MLTQParser’s qf field specifies the fields to use for similarity. We will require amintf (Minimum Term Frequency) of 1. This is the frequency below which search terms will be ignored in the source document. We will also require a mindf (Minimum Document Frequency) of 1. This is the frequency at which words will be ignored when they do not occur in at least this many documents.

The results of the Solr search appear logical, they are the other five Star Wars movies. Note that since the first five results are exact matches on The Phantom Menace’s three genres, ‘action’, ‘adventure’, and ‘fantasy’, their scores are identical, 6.33.

Solr:

Parameters
----------
q: {!mlt qf="genres" mintf=1 mindf=1}aaf956a5-afa5-4284-91c1-69455142884f
kwargs: {
  'defType': 'lucene', 
  'fq': 'countries: USA', 
  'fl': 'title, genres, score', 
  'rows': '5'
}
  
Results
----------
document count: 252
{'title': 'Star Wars: Episode IV - A New Hope', 'genres': ['Action', 'Adventure', 'Fantasy'], 'score': 6.33}
{'title': 'Star Wars: Episode V - The Empire Strikes Back', 'genres': ['Action', 'Adventure', 'Fantasy'], 'score': 6.33}
{'title': 'Star Wars: Episode VI - Return of the Jedi', 'genres': ['Action', 'Adventure', 'Fantasy'], 'score': 6.33}
{'title': 'Star Wars: Episode III - Revenge of the Sith', 'genres': ['Action', 'Adventure', 'Fantasy'], 'score': 6.33}
{'title': 'Star Wars: Episode II - Attack of the Clones', 'genres': ['Action', 'Adventure', 'Fantasy'], 'score': 6.33}

Here is the actual Lucene query Solr will run.

+(genres:action genres:adventur genres:fantasi) 
-id:652adcfa-c59c-4fa0-ace5-6345fec3cfff

Query 9b: The Problem with George

In the second MLTQParser query example, we will again choose the movie, ‘Star Wars: Episode I – The Phantom Menace’. However, this time will look for similar movies based on a comparison of the actors, director, and writers fields (shown below). Basically, we are looking for similarities between the people associated with ‘Star Wars: Episode I – The Phantom Menace’ and other movies.

Solr:

Parameters
----------
q: id:"aaf956a5-afa5-4284-91c1-69455142884f"
kwargs: {
  'defType': 'lucene', 
  'fl': 'actors, director, writers'
}
  
Results
----------
document count: 1
{
  'director': ['George Lucas'], 
  'writers': ['George Lucas'], 
  'actors': ['Liam Neeson', 'Ewan McGregor', 'Natalie Portman', 'Jake Lloyd']
}

As you can tell by the results of the MLTQParser query below, the first nine out of ten search results make sense. However, the tenth movie result, ‘New Meet Me on South Street: The Story of JC Dobbs’, has no obvious similarities with ‘Star Wars: Episode I – The Phantom Menace’. The movie does not share a common director, writer, or actor.

Solr:

Parameters
----------
q: {!mlt qf="actors director writers" mintf=1 mindf=1}aaf956a5-afa5-4284-91c1-69455142884f
kwargs: {
  'defType': 'lucene', 
  'fq': 'countries: USA', 
  'fl': 'title, actors, director, writers, score', 
  'rows': '10'
}
  
Results
----------
document count: 55
{'title': 'Star Wars: Episode III - Revenge of the Sith', 'director': ['George Lucas'], 'writers': ['George Lucas'], 'actors': ['Ewan McGregor', 'Natalie Portman', 'Hayden Christensen', 'Ian McDiarmid'], 'score': 44.84}
{'title': 'Star Wars: Episode II - Attack of the Clones', 'director': ['George Lucas'], 'writers': ['George Lucas', 'Jonathan Hales', 'George Lucas'], 'actors': ['Ewan McGregor', 'Natalie Portman', 'Hayden Christensen', 'Christopher Lee'], 'score': 44.58}
{'title': 'Star Wars: Episode IV - A New Hope', 'director': ['George Lucas'], 'writers': ['George Lucas'], 'actors': ['Mark Hamill', 'Harrison Ford', 'Carrie Fisher', 'Peter Cushing'], 'score': 23.51}
{'title': 'Star Wars: Episode VI - Return of the Jedi', 'director': ['Richard Marquand'], 'writers': ['Lawrence Kasdan', 'George Lucas', 'George Lucas'], 'actors': ['Mark Hamill', 'Harrison Ford', 'Carrie Fisher', 'Billy Dee Williams'], 'score': 11.96}
{'title': 'A Million Ways to Die in the West', 'director': ['Seth MacFarlane'], 'writers': ['Seth MacFarlane', 'Alec Sulkin', 'Wellesley Wild'], 'actors': ['Seth MacFarlane', 'Charlize Theron', 'Amanda Seyfried', 'Liam Neeson'], 'score': 11.72}
{'title': 'Run All Night', 'director': ['Jaume Collet-Serra'], 'writers': ['Brad Ingelsby'], 'actors': ['Liam Neeson', 'Ed Harris', 'Joel Kinnaman', 'Boyd Holbrook'], 'score': 11.72}
{'title': 'I Love You Phillip Morris', 'director': ['Glenn Ficarra, John Requa'], 'writers': ['John Requa', 'Glenn Ficarra', 'Steve McVicker'], 'actors': ['Jim Carrey', 'Ewan McGregor', 'Leslie Mann', 'Rodrigo Santoro'], 'score': 10.97}
{'title': 'The Island', 'director': ['Michael Bay'], 'writers': ['Caspian Tredwell-Owen', 'Alex Kurtzman', 'Roberto Orci', 'Caspian Tredwell-Owen'], 'actors': ['Ewan McGregor', 'Scarlett Johansson', 'Djimon Hounsou', 'Sean Bean'], 'score': 10.97}
{'title': 'Big Fish', 'director': ['Tim Burton'], 'writers': ['Daniel Wallace', 'John August'], 'actors': ['Ewan McGregor', 'Albert Finney', 'Billy Crudup', 'Jessica Lange'], 'score': 10.97}
{'title': 'New Meet Me on South Street: The Story of JC Dobbs', 'director': ['George Manney'], 'writers': ['George Manney'], 'actors': ['Tony Bidgood', 'Peter Stone Brown', 'Stephen Caldwell', 'Tommy Conwell'], 'score': 10.5}

Here is the actual Lucene query Solr run.

+(writers:george director:george actors:natalie writers:lucas 
  actors:jake actors:portman actors:ewan actors:mcgregor 
  actors:liam actors:lloyd director:lucas actors:neeson) 
-id:652adcfa-c59c-4fa0-ace5-6345fec3cfff

Examining the query, we can plainly see the problem with MLTQParser. The MLTQParser query is splitting the first and last names of actors, directors, and writers, then searching for each name individually, but not their whole name. In my opinion, this is a bug with the MLTQParser, since each value in the actors, director, and writers MultiValued fields are wrapped in double quotes. The query should treat each value an exact phrase, not individual search terms.

Given the MLTQParser’s query logic, it is now clear why a seemingly irrelevant movie, like ‘New Meet Me on South Street: The Story of JC Dobbs’, was part of the search results. Examining the debug output of the scoring explanation, we see the following.

"544637e5-e96d-4f62-9b22-5174c60ee512":"
10.504457 = sum of:
  10.504457 = sum of:
    5.5608187 = weight(writers:george in 649) [SchemaSimilarity], result of:
      5.5608187 = score(doc=649,freq=1.0 = termFreq=1.0
), product of:
        4.226696 = idf, computed as log(1 + (docCount - docFreq + 0.5) / (docFreq + 0.5)) from:
          26.0 = docFreq
          1814.0 = docCount
        1.315642 = tfNorm, computed as (freq * (k1 + 1)) / (freq + k1 * (1 - b + b * fieldLength / avgFieldLength)) from:
          1.0 = termFreq=1.0
          1.2 = parameter k1
          0.75 = parameter b
          4.836273 = avgFieldLength
          2.0 = fieldLength
    4.943639 = weight(director:george in 649) [SchemaSimilarity], result of:
      4.943639 = score(doc=649,freq=1.0 = termFreq=1.0
), product of:
        4.6206594 = idf, computed as log(1 + (docCount - docFreq + 0.5) / (docFreq + 0.5)) from:
          20.0 = docFreq
          2081.0 = docCount
        1.069899 = tfNorm, computed as (freq * (k1 + 1)) / (freq + k1 * (1 - b + b * fieldLength / avgFieldLength)) from:
          1.0 = termFreq=1.0
          1.2 = parameter k1
          0.75 = parameter b
          2.3801057 = avgFieldLength
          2.0 = fieldLength
"

George Manney was both the Director and Writer of ‘New Meet Me on South Street: The Story of JC Dobbs’. George Manney shares a first name with George Lucas, the Director and Writer of ‘Star Wars: Episode I – The Phantom Menace’. This is the only similarity between people associated with both movies. Therefore, there were two matches, one match on the director fields and a match on the writers field. Unfortunately, having the same first names negatively impacts the results from the MLTQParser.

Synonyms

For our last example, we will examine the use of synonyms with Solr. In respect to the movie index, we could perform the following eDisMax search on the title, plot, and genres fields: ‘scary’ OR ‘slasher’ OR ‘spooky’ OR ‘evil’ OR ‘horror’, or more simply ‘scary slasher spooky evil horror’. Based on this search, we would get back a truly gruesome collection of 141 films. The search is effective because it uses multiple, similar search terms to return a larger resultset of movies within a similar theme. However, the search relies on each end-user to enter the same relevant search terms, every time.

With Solr’s synonym capability, we can build some intelligence into our index by defining synonymous relationships between terms. There are multiple ways to define synonymous relationships between terms in Solr. Lucidworks has an excellent article, Synonyms Files, on the different synonymous relationships. We will look at three types of relationships, as described by Lucidworks: Replacement Synonyms, Oneway Expansion Synonyms, and Multiway Expansion Synonyms.

I have pre-define some examples for each of the three types of relationships, in the movies index’s synonmys.txt configuration file. This file is created when a new index is created.

## Custom synonym groups for movies index ##

# Replacement Synonyms examples
scarey => scary
ciborg => cyborg

# Multiway Expansion Synonyms examples
scary,slasher,spooky,evil,horror

# Oneway Expansion Synonyms examples
droid => droid,android,robot,cyborg

There is a copy of the file in this project, which will be used by the movies index, running in the Docker container.

Query 10a: Replacement Synonyms

We will start with a Replacement Synonyms example. I have added the following synonym mapping for a common misspelling of the word ‘cyborg’.

ciborg => cyborg

If we perform a search for the term ‘ciborg’, Solr will substitute it with the term ‘cyborg’. We can confirm this by viewing the query, as shown in the debug snippet below.

+(title:cyborg | plot:cyborg | genres:cyborg)

Performing the query returns two documents, including the most famous cyborg, Arnold Schwarzenegger, the Terminator.

Parameters
----------
q: ciborg
kwargs: {
  'defType': 'edismax', 
  'qf': 'title plot genres', 
  'fl': 'title, score', 
  'rows': '5'
}

Results
----------
document count: 2
{'title': 'Terminator 2: Judgment Day', 'score': 8.17}
{'title': "I'm a Cyborg, But That's OK", 'score': 7.13}

Query 10b: Oneway Expansion Synonyms

Next, we will demonstrate Oneway Expansion Synonyms. I have added the following synonym mapping for the word, ‘droid’. Note we must include the word ‘droid’ in the expansion synonyms on the right side of the mapping, as well as on the left.

droid => droid,android,robot,cyborg

When we perform a search on the term ‘droid’, Solr will also search for the synonyms ‘android’, ‘robot’, and ‘cyborg’. We can confirm this by viewing the query Solr constructs for the search term ‘droid’, as shown in the debugger snippet below.

+(
  Synonym(title:android title:cyborg title:droid title:robot) | 
  Synonym(plot:android plot:cyborg plot:droid plot:robot) | 
  Synonym(genres:android genres:cyborg genres:droid genres:robot)
)

Note the converse is not true since this is a one-way relationship. If we search on ‘cyborg’, Solr will not search on ‘droid’, ‘android’, and ‘robot’.

+(title:cyborg | plot:cyborg | genres:cyborg)

Performing the ‘droid’ eDisMax query returns 15 documents.

Parameters
----------
q: droid
kwargs: {
  'defType': 'edismax', 
  'qf': 'title plot genres', 
  'fl': 'title, score', 
  'rows': '5'
}

Results
----------
document count: 15
{'title': 'Robo Jî', 'score': 7.67}
{'title': "I'm a Cyborg, But That's OK", 'score': 7.13}
{'title': 'BV-01', 'score': 6.6}
{'title': 'Robot Chicken: DC Comics Special', 'score': 6.44}
{'title': 'Terminator 2: Judgment Day', 'score': 6.23}

Query 10c: Multiway Expansion Synonyms

Lastly, we will demonstrate Multiway Expansion Synonyms. I have added the following synonym mapping for the word, ‘scary’.

scary,slasher,spooky,evil,horror

If we perform a search on the term ‘scary’, Solr will also search for the synonyms ‘slasher’, ‘spooky’, ‘evil’, and ‘horror’. Unlike the previous example, the converse is true since this is a multi-way relationship. If we search on any of the five synonyms, Solr will also search on the other four terms and return identical results. We can confirm this by viewing the query Solr constructs for the term ‘scary’, as shown in the debug snippet below.

+(Synonym(title:evil title:horror title:scari title:slasher 
  title:spooki) | Synonym(plot:evil plot:horror plot:scari 
  plot:slasher plot:spooki) | Synonym(genres:evil genres:horror 
  genres:scari genres:slasher genres:spooki))

Performing the ‘scary’ eDisMax query returns 141 documents.

Parameters
----------
q: scary
kwargs: {
  'defType': 'edismax', 
  'qf': 'title plot genres', 
  'fl': 'title, score', 
  'rows': '5'
}

Results
----------
document count: 141
{'title': 'See No Evil, Hear No Evil', 'score': 7.9}
{'title': 'The Evil Dead', 'score': 7.23}
{'title': 'Evil Dead', 'score': 7.23}
{'title': 'Evil Ed', 'score': 7.23}
{'title': 'Evil Dead II', 'score': 6.37}

You may have noticed the other replacement synonym mapping I placed in the index’s synonmys.txt configuration file, shown below.

scarey => scary

You might be wondering if you entered the common misspelling ‘scarey’ as a search term, would Solr replace the term with ‘scary’ and search on the term ‘scary’, but also search on it’s four synonyms, ‘slasher’, ‘spooky’, ‘evil’, and ‘horror’. The answer is no, Solr will only search on ‘scary’. However, Solr does not create indirect or secondary relationships between synonym mappings. You would have to correct the spelling and input the term correctly to take advantage of the multi-way relationship.

Managing Unique Vocabulary with Synonyms

Large corporations, industry verticles, government agencies, and other entities often use a unique lexicon. Their vocabulary includes uncommon terms, phrases, acronyms, abbreviations, and other idioms. These commonly represent products and services, organizational structures, and technical jargon. Synonyms are an excellent way to support domain-specific dictions within indexes. We see this in the examples Solr includes in the synonmys.txt configuration file. The two examples, shown below, demonstrate how associate multiple abbreviations and technical terms to equivalent amounts of compute.

# Some synonym groups specific to this example
GB,gib,gigabyte,gigabytes
MB,mib,megabyte,megabytes

10d: Synonymous Phrases

A largely undocumented feature of synonyms is the ability to create synonymous relationships between common acronyms, abbreviations, terms, and phrases. Below I have provided a few examples of how to create these relationships. The only apparent, logical limitation, you cannot use stop words in the phrases; a good reason not to overuse stop words.

ai,artificial intelligence
cia,central intelligence agency
fbi,federal bureau investigation
lol,laughing out loud,league legends

For example, If we include the full phrase ‘league of legends’ in the synonyms map with ‘lol’, Solr ignores this phrase when I search on ‘LOL’. However, if we remove the stop word, ‘of’, then Solr will create a query that requires the two terms, ‘league’ and ‘legends’, or the two terms ‘laugh’ and ‘loud’, or the single term ‘lol’.

+(
  (((+title:laugh +title:out +title:loud) (+title:leagu +title:legend) title:lol)) | 
  (((+plot:laugh +plot:out +plot:loud) (+plot:leagu +plot:legend) plot:lol)) | 
  (((+genres:laugh +genres:out +genres:loud) (+genres:leagu +genres:legend) genres:lol))
)

Solr:

Parameters
----------
q: lol
kwargs: {
  'defType': 'edismax', 
  'qf': 'title plot genres', 
  'fl': 'title, score', 
  'rows': '5'
}

Results
----------
document count: 1
{'title': 'JK LOL', 'score': 9.05}

Conclusion

By understanding the capabilities of Apache Solr, the characteristics of the data contained in your indexes and the search patterns of your end users, you will be able to craft queries that ensure responses contain high-quality, relevant search results.

The query examples in this post demonstrate only a very small portion of Solr’s vast search capabilities. There are several additional query examples available in each of the two Python scripts, which you can uncomment and explore their results, further.

Solr’s documentation is very good; to learn more about Solr’s capabilities, I suggest reviewing the various parsers and their options, in the current Solr version 7.6 documentation.

• The Standard Query Parser
• The DisMax Query Parser
• The Extended DisMax Query Parser
• Other Parsers

I also suggest reviewing Solr’s Analyzers, Tokenizers, and Filters, and understand how they affect the way in which Solr indexes documents and how Solr interprets the content of the indexes when performing a search.

• Analyzers, Tokenizers, and Filters
• Tokenizers
• Filter Descriptions
• Language Analysis

 

All opinions expressed in this post are my own and not necessarily the views of my current or past employers or their clients.

Feature Illustration Copyright: Dejan Bozic (123RF)