SlideShare a Scribd company logo

Encores

OpenSource Connections, profile picture
OpenSource Connections

This document discusses various search features beyond basic matching and ranking, including facets, query auto-completion, spelling correction, and query relaxation. It provides examples of how these features are implemented in Solr to help users formulate queries, understand results, and narrow down search outputs. Specific challenges with facets for e-commerce search are outlined, such as handling product variants and selecting the best facet values. Solutions proposed include indexing one document per variant, using collapse queries, and executing a prior facet request to select meaningful facets. The document also discusses approaches to auto-completion using suggesters and spelling correction using the Solr spellcheck component or alternative methods like fuzzy search over a query index. Finally, query relaxation techniques are briefly covered, such

Technology
1 of 53
Download to read offline
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
Encores? - Going beyond matching and ranking of search results Berlin Buzzwords 2021 Eric Pugh, René Kriegler
Who we are @renekrie @dep4b Combined 30 years of experience in search Open Source enthusiasts ASF member, Committers on: Solr, Querqy, SMUI, Quepid,
https://mices.co
Search - beyond matching and ranking We tend to focus on matching and ranking. Other search features are almost treated like an afterthought, like ‘encores’ that follow the main performance: ● Facets ● Query auto-completion ● Spelling correction ● Query relaxation => BUT: These are essential features that help the user formulate the query, understand and narrow down the results
Our main act today (not encores!) ● Facets ● Query auto-completion ● Spelling correction ● Query relaxation Learn about solutions that come out of the box (in Solr) Typical challenges and how to overcome them Advanced solutions: understand the concepts, create your own
The Art of Facets
Facets help the user ... ● understand the search results (see ‘what is there’, learn about the domain) ● narrow down search results Chorus Electronics Project: https://github.com/querqy/chorus Try the Demo Ecommerce Shop: http://chorus.dev.o19s.com:4000/
Facets help the user ... ● understand the search results (see ‘what is there’, learn about the domain) ● narrow down search results
Challenges Getting the counts right in e-commerce search Showing the best facets in the best order Selecting the facet values to show
“Qui numerare incipit errare incipit” Facet Counts
Facets and filters A trivial example: query=t-shirts filter=color:black Challenge: We still need to count all colours in the facets, even if the search result contains only black t-shirts Solution: Tagging and exclusion of filters
Facets and filters: tagging and exclusion Tagging: fq={!tag=f_color}color:black Exclusion: Facet param facet.field={!ex=f_color}color JSON facets "facet": { "color": { "type": "terms", "field": "color", "domain": { "excludeTags":"f_color" } } }
Challenge: product variants Product ID: 9739, brand: “inteemate” Size: XS Price: 11.99 Size: XL Price: 11.99 Size: S Price: 12.99 Size: S Price: 12.99 Size: M Price: 13.99 Size: L Price: 13.99
Challenge: product variants Product ID: 9739, brand: “inteemate” Size: XS Price: 11.99 Size: XL Price: 11.99 Size: S Price: 12.99 Size: S Price: 12.99 Size: M Price: 13.99 Size: L Price: 13.99 color: [green, yellow, blue] size: [XS, S, M, L, XL] price: [11.99, 12.99, 13.99] Merge into single document?? Facets would work great but false matches for filter color:green AND size:M
Challenge: product variants Best solution (in our opinion): ● Index one document for each variant ● Group variants at query time using the collapse query parser: fq={!collapse field=productId} => Boolean filters work as expected => Great flexibility for counting facets => Fast enough
Challenge: product variants Size: XS Price: 11.99 Product: 9739 Brand: inteemate Size: XL Price: 11.99 Product: 9739 Brand: inteemate Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: M Price: 13.99 Product: 9739 Brand: inteemate Size: L Price: 13.99 Product: 9739 Brand: inteemate filter query fq=brand:inteemate
Challenge: product variants Size: XS Price: 11.99 Product: 9739 Brand: inteemate Size: XL Price: 11.99 Product: 9739 Brand: inteemate Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: M Price: 13.99 Product: 9739 Brand: inteemate Size: L Price: 13.99 Product: 9739 Brand: inteemate filter query fq=brand:inteemate filter query fq=color:blue
Challenge: product variants Size: XS Price: 11.99 Product: 9739 Brand: inteemate Size: XL Price: 11.99 Product: 9739 Brand: inteemate Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: M Price: 13.99 Product: 9739 Brand: inteemate Size: L Price: 13.99 Product: 9739 Brand: inteemate filter query fq=brand:inteemate filter query fq=color:blue query q=t-shirt
Challenge: product variants Size: XS Price: 11.99 Product: 9739 Brand: inteemate Size: XL Price: 11.99 Product: 9739 Brand: inteemate Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: M Price: 13.99 Product: 9739 Brand: inteemate Size: L Price: 13.99 Product: 9739 Brand: inteemate filter query fq=brand:inteemate filter query fq=color:blue query q=t-shirt post filter query fq={!collapse field=productId}
Challenge: product variants in facets Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: M Price: 13.99 Product: 9739 Brand: inteemate Size: L Price: 13.99 Product: 9739 Brand: inteemate ... post filter query fq={!collapse field=productId} Facet counts will be correct for product attributes (“brand”)
Challenge: product variants in facets Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: M Price: 13.99 Product: 9739 Brand: inteemate Size: L Price: 13.99 Product: 9739 Brand: inteemate ... post filter query fq={!collapse field=productId tag=coll} For facet counts of variant attributes we’ll have to tag and exclude collapse filter: "facet": { "size": { "type": "terms", "field": "size", "domain": { "excludeTags":"coll" } } } Sizes S, M, L all shown as ‘1 result’ in facets ✅
Challenge: product variants in facets Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: M Price: 13.99 Product: 9739 Brand: inteemate Size: L Price: 13.99 Product: 9739 Brand: inteemate ... post filter query fq={!collapse field=productId tag=coll} For facet counts of variant attributes we’ll have to tag and exclude collapse filter: "facet": { "color": { "type": "terms", "field": "color", "domain": { "excludeTags":"coll" } } } Sizes S, M, L all shown as ‘1 result’ in facets ✅ Color ‘Blue’ shown as ‘3 results’ in facets ❌
Challenge: product variants in facets Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: M Price: 13.99 Product: 9739 Brand: inteemate Size: L Price: 13.99 Product: 9739 Brand: inteemate ... post filter query fq={!collapse field=productId tag=coll} For facet counts of variant attributes we’ll have to tag and exclude collapse filter: "facet": { "color": { "type": "terms", "field": "color", "domain": { "excludeTags":"coll" }, facet: { "numProducts":"unique(productId)" } } } Sizes S, M, L all shown as ‘1 result’ in facets ✅ Color ‘Blue’ shown as ‘1 result’ in facets ✅
Collapse query parser - notes on implementation ● Beware of high cardinality of product IDs. ○ If you have 10M different product IDs in your index, the collapse query parser will allocate heap space for 2 arrays (float/int) x 10M elements (ca. 80 MB) per request! ○ Solution: ■ Many products have just 1 variant. It’s better to leave the productId empty in this case. ■ Combine with nullPolicy=expand, which avoids reserving array space for products without a productId: fq={!collapse field=productId nullPolicy=expand} ● All variants of a product must be indexed to the same shard
Facet Selection
Which facets should we show? Some domains are rich in attributes. For example, electronics could use 10k different attributes. Even if we reduced the number of attributes to be used in facets at index time, we could be left with several hundreds of candidates for facetting. Building a request for hundreds of facets is not feasible. We’ll show a simple solution, that will just use the search engine to select facets. At the other end of the spectrum, you could train a model, that predicts which facets to show for a given query.
Which facets should we show? - Solution Index a field multivalued field that holds the names of the facettable fields... Doc1: screenSize: 17, .... facettableFields: [ “screenSize”, “ramGB”, “height”, “width”, ... ] Doc2: ... facettableFields: [ “screenSize”, “numHDMIPorts”, “height”, “width”, ... ]
Which facets should we show? - Solution ... and execute an additional, prior facet request on this field. Add the facet values returned by this request as facet parameters to the main request: "facet": { "facettable_fields": { "type": "terms", "field": "facettable_fields" } } (query/filter queries are the same like in the ‘main request’) "facets":{ "facettable_fields":{ "buckets":[{ "val":"screenSize", "count":12}, { "val":"ramGB", "count":4}, { "val":"height", "count":3}, ] } }
Which facets should we show? - Solution Index additional information together with the names of facettable fields facettableFields: [ "00010;screenSize;Screen size ", "00100;ramGB;Memory (GB) ", "00005;height;Height ", "00005;width;Width ", ...] Importance (padding makes values sortable!) Field name Label
Facet Value Selection
Which facet values? Category Pills being dynamically included IF the entropy model says they are meaningful for filtering the data
Shannon’s Entropy Worksheet https://bit.ly/measure-diversity (https://en.wikipedia.org/wiki/Entropy_(information_theory))
Auto-completion & spelling correction
Autocompletion - Using a Suggester <searchComponent name="suggest" class="solr.SuggestComponent"> <lst name="suggester"> <str name="name">mySuggester</str> <str name="lookupImpl">FuzzyLookupFactory</str> <str name="suggestAnalyzerFieldType">text_general</str> <str name="buildOnCommit">true</str> <str name="field">dictionary</str> </lst> </searchComponent> Experiment with combinations of Lookups & Dictionary implementations.
Spellchecking - Using Solr component Two flavours: “cofffee --> coffee”, collations: “expresso machine” -->“espresso machine” <str name="spellcheck">true</str> <str name="spellcheck.dictionary">title</str> <str name="spellcheck.onlyMorePopular">true</str> <str name="spellcheck.extendedResults">true</str> <str name="spellcheck.collate">true</str> <str name="spellcheck.maxCollations">100</str> <str name="spellcheck.maxCollationTries">5</str> <str name="spellcheck.count">5</str> <str name="spellcheck.collateParam.mm">100%</str> <searchComponent name="suggest" class="solr.SuggestComponent"> <lst name="suggester"> <str name="name">mySuggester</str> <str name="lookupImpl">FuzzyLookupFactory</str> <str name="suggestAnalyzerFieldType">text_general</str> <str name="buildOnCommit">true</str> <str name="field">dictionary</str> </lst> </searchComponent> Good collations are what we want!
Autocompletion - using a query index User enters: ja Show the best query completions in the best order! Using ideas from Joshua Bacher/Christine Bellstedt, Search Suggestions - The Underestimated Killer Feature of your Online Shop. Berlin Buzzwords 2018
Autocompletion - using a query index User enters: ja Match prefix in field “match” q=*:*&fq=match:ja indexed with EdgeNGramFilter, lowercase, remove accents/ASCII folding, ... Optionally index and match spelling variants (jacket/jakcet)
Autocompletion - using a query index Sort by “weight desc” q=*:*&fq=match:ja &sort=weight desc Sorting might get slow for short prefixes if the query index is large - tag the top N queries for lengths 1 and 2 and add another filter (fewer matches to sort, nicely cacheable): q=*:*&fq=match:ja &sort=weight desc &fq=top_len_2:true
Autocompletion - using a query index If two queries have the same fingerprint, drop the one with the lower weight Fingerprint: concatenated sorted, normalised query tokens This increases the diversity of the suggestions.
Autocompletion - using a query index Suggest the Labels as query completions!
Autocompletion - using a query index You can show the best matching category for disambiguation and affirmation: * jacket * jacket in Fashion
Spelling correction - using a query index Structure similar to query index for autocompletion Copy of the ‘match’ field indexed as n-grams
Spelling correction - using a query index Filters on edit distance and rank based on n-grams (via TF*IDF) q=jakc jones &defType=edismax &qf=match_ngram &sow=false &fq=match:jakc jones~2
Spelling correction - using a query index Add boost by weight (or a function of it) q=jakc jones &defType=edismax &qf=match_ngram &sow=false &fq=match:jakc jones~2 &boost=weight
General model for spelling correction & autocompletion Noisy Channel Model / Bayesian Inference (Kernighan et. al., 1990; Jurafsky & Martin, 2009) Our ‘Weight’ field Edit distance, n-gram model, keyboard layout (Symspell!), ... prefix match for autocompletion
Query relaxation
Query relaxation Which query term should we drop if we can’t match all of them together? jacket xs green jacket xs green jacket xs green jacket xs green iphone 12 iphone 12 iphone 12
Query relaxation - ‘mm’ anti-pattern Loosening ‘minimum should match’ (mm) constraint to < 100% iphone 12 iphone 12 You’ll get matches for “12” She will just see probably imprecise results that don’t match her query exactly. You cannot tell the user what happened and which term you dropped. She wouldn’t know what to do in order to improve the query. Don’t do this! At least not in e-commerce search
Query relaxation - Solutions René.Kriegler, Query Relaxation - a rewriting technique between search and recommendations. Haystack Conference 2019
Query relaxation - Solutions Try searching with each term individually and drop the one from the query that yields the fewest results (might require additional rules to avoid just keeping number terms)
Query relaxation - Solutions Multi-layer Neural Network, Word embeddings as input to represent terms
Encores? Facets, autocompletion, spelling correction, query relaxation are important features of a search application. We’ve shown simple out-of-the-box solutions and a path to implement more advanced approaches.
Thank you!

More Related Content

PPT
Modern information Retrieval-Relevance Feedback
HasanulFahmi2
 
PPTX
ORC File & Vectorization - Improving Hive Data Storage and Query Performance
DataWorks Summit
 
PDF
Faceted Search with Lucene
lucenerevolution
 
PDF
Building a Complex, Real-Time Data Management Application
Jonathan Katz
 
PPTX
LinkedIn talk at Netflix ML Platform meetup Sep 2019
Faisal Siddiqi
 
PDF
Let's Build an Inverted Index: Introduction to Apache Lucene/Solr
Sease
 
PPTX
Automatically Build Solr Synonyms List using Machine Learning - Chao Han, Luc...
Lucidworks
 
PDF
Geotermia a Bassa Entalpia
Felicetto Massa
 
Modern information Retrieval-Relevance Feedback
HasanulFahmi2
 
ORC File & Vectorization - Improving Hive Data Storage and Query Performance
DataWorks Summit
 
Faceted Search with Lucene
lucenerevolution
 
Building a Complex, Real-Time Data Management Application
Jonathan Katz
 
LinkedIn talk at Netflix ML Platform meetup Sep 2019
Faisal Siddiqi
 
Let's Build an Inverted Index: Introduction to Apache Lucene/Solr
Sease
 
Automatically Build Solr Synonyms List using Machine Learning - Chao Han, Luc...
Lucidworks
 
Geotermia a Bassa Entalpia
Felicetto Massa
 

What's hot (10)

PDF
Encodings
Nicola Pastorello, Ph.D.
 
PDF
Facebook Talk at Netflix ML Platform meetup Sep 2019
Faisal Siddiqi
 
PDF
Cardinality Estimation through Histogram in Apache Spark 2.3 with Ron Hu and ...
Databricks
 
PDF
Netflix Global Search - Lucene Revolution
ivan provalov
 
PDF
Dense Retrieval with Apache Solr Neural Search.pdf
Sease
 
PDF
Data Security at Scale through Spark and Parquet Encryption
Databricks
 
PDF
Incremental View Maintenance with Coral, DBT, and Iceberg
Walaa Eldin Moustafa
 
PDF
[2A1]Line은 어떻게 글로벌 메신저 플랫폼이 되었는가
NAVER D2
 
PDF
Correlation, causation and incrementally recommendation problems at netflix ...
Roelof van Zwol
 
PDF
Tutorial on Deep Learning in Recommender System, Lars summer school 2019
Anoop Deoras
 
Encodings
Nicola Pastorello, Ph.D.
 
Facebook Talk at Netflix ML Platform meetup Sep 2019
Faisal Siddiqi
 
Cardinality Estimation through Histogram in Apache Spark 2.3 with Ron Hu and ...
Databricks
 
Netflix Global Search - Lucene Revolution
ivan provalov
 
Dense Retrieval with Apache Solr Neural Search.pdf
Sease
 
Data Security at Scale through Spark and Parquet Encryption
Databricks
 
Incremental View Maintenance with Coral, DBT, and Iceberg
Walaa Eldin Moustafa
 
[2A1]Line은 어떻게 글로벌 메신저 플랫폼이 되었는가
NAVER D2
 
Correlation, causation and incrementally recommendation problems at netflix ...
Roelof van Zwol
 
Tutorial on Deep Learning in Recommender System, Lars summer school 2019
Anoop Deoras
 
Ad

Similar to Encores (20)

PDF
Faceted Search And Result Reordering
Varun Thacker
 
PDF
Search refinement
Neoworks
 
PDF
The Many Facets of Apache Solr - Yonik Seeley
lucenerevolution
 
PDF
Real-Time Analytics with Solr: Presented by Yonik Seeley, Cloudera
Lucidworks
 
PDF
Automatically mining facets for queries from their search results
Shakas Technologies
 
PDF
A Survey on Automatically Mining Facets for Queries from their Search Results
IRJET Journal
 
PDF
Search@flipkart
Umesh Prasad
 
PDF
Query Recommendation by using Collaborative Filtering Approach
IRJET Journal
 
PDF
Finding Love with MongoDB
MongoDB
 
PDF
How to Build your Training Set for a Learning To Rank Project - Haystack
Sease
 
DOCX
GENERATING QUERY FACETS USING KNOWLEDGE BASES
Prasadu Peddi
 
PDF
Solr 3.1 and beyond
Lucidworks (Archived)
 
PPT
What to do when one size does not fit all?!
Arjen de Vries
 
PDF
elasticsearch - advanced features in practice
Jano Suchal
 
PDF
Smart Facets at Rakuten: Presented by Keith Thoma & Michael Pellegrini, Rakut...
Lucidworks
 
PDF
Parallel SQL and Analytics with Solr: Presented by Yonik Seeley, Cloudera
Lucidworks
 
PDF
Fulltext engine for non fulltext searches
Adrian Nuta
 
PPTX
Lots of facets, fast
BeyondTrees
 
PDF
Fast & relevant search: solutions and trade-offs (January 2020 - Search Techn...
Sylvain Utard
 
PPTX
Effective and Efficient Entity Search in RDF data
Roi Blanco
 
Faceted Search And Result Reordering
Varun Thacker
 
Search refinement
Neoworks
 
The Many Facets of Apache Solr - Yonik Seeley
lucenerevolution
 
Real-Time Analytics with Solr: Presented by Yonik Seeley, Cloudera
Lucidworks
 
Automatically mining facets for queries from their search results
Shakas Technologies
 
A Survey on Automatically Mining Facets for Queries from their Search Results
IRJET Journal
 
Search@flipkart
Umesh Prasad
 
Query Recommendation by using Collaborative Filtering Approach
IRJET Journal
 
Finding Love with MongoDB
MongoDB
 
How to Build your Training Set for a Learning To Rank Project - Haystack
Sease
 
GENERATING QUERY FACETS USING KNOWLEDGE BASES
Prasadu Peddi
 
Solr 3.1 and beyond
Lucidworks (Archived)
 
What to do when one size does not fit all?!
Arjen de Vries
 
elasticsearch - advanced features in practice
Jano Suchal
 
Smart Facets at Rakuten: Presented by Keith Thoma & Michael Pellegrini, Rakut...
Lucidworks
 
Parallel SQL and Analytics with Solr: Presented by Yonik Seeley, Cloudera
Lucidworks
 
Fulltext engine for non fulltext searches
Adrian Nuta
 
Lots of facets, fast
BeyondTrees
 
Fast & relevant search: solutions and trade-offs (January 2020 - Search Techn...
Sylvain Utard
 
Effective and Efficient Entity Search in RDF data
Roi Blanco
 
Ad

More from OpenSource Connections (20)

PDF
Why User Behavior Insights? KMWorld Enterprise Search & Discovery 2024
OpenSource Connections
 
PDF
Test driven relevancy
OpenSource Connections
 
PDF
How To Structure Your Search Team for Success
OpenSource Connections
 
PPT
The right path to making search relevant - Taxonomy Bootcamp London 2019
OpenSource Connections
 
PDF
Payloads and OCR with Solr
OpenSource Connections
 
PPTX
Haystack 2019 Lightning Talk - The Future of Quepid - Charlie Hull
OpenSource Connections
 
PDF
Haystack 2019 Lightning Talk - State of Apache Tika - Tim Allison
OpenSource Connections
 
PPTX
Haystack 2019 Lightning Talk - Relevance on 17 million full text documents - ...
OpenSource Connections
 
PPTX
Haystack 2019 Lightning Talk - Solr Cloud on Kubernetes - Manoj Bharadwaj
OpenSource Connections
 
PDF
Haystack 2019 Lightning Talk - Quaerite a Search relevance evaluation toolkit...
OpenSource Connections
 
PPTX
Haystack 2019 - Search-based recommendations at Politico - Ryan Kohl
OpenSource Connections
 
PPTX
Haystack 2019 - Search with Vectors - Simon Hughes
OpenSource Connections
 
PPTX
Haystack 2019 - Natural Language Search with Knowledge Graphs - Trey Grainger
OpenSource Connections
 
PPTX
Haystack 2019 - Search Logs + Machine Learning = Auto-Tagging Inventory - Joh...
OpenSource Connections
 
PDF
Haystack 2019 - Improving Search Relevance with Numeric Features in Elasticse...
OpenSource Connections
 
PDF
Haystack 2019 - Architectural considerations on search relevancy in the conte...
OpenSource Connections
 
PPTX
Haystack 2019 - Custom Solr Query Parser Design Option, and Pros & Cons - Ber...
OpenSource Connections
 
PPTX
Haystack 2019 - Establishing a relevance focused culture in a large organizat...
OpenSource Connections
 
PPTX
Haystack 2019 - Solving for Satisfaction: Introduction to Click Models - Eliz...
OpenSource Connections
 
PDF
2019 Haystack - How The New York Times Tackles Relevance - Jeremiah Via
OpenSource Connections
 
Why User Behavior Insights? KMWorld Enterprise Search & Discovery 2024
OpenSource Connections
 
Test driven relevancy
OpenSource Connections
 
How To Structure Your Search Team for Success
OpenSource Connections
 
The right path to making search relevant - Taxonomy Bootcamp London 2019
OpenSource Connections
 
Payloads and OCR with Solr
OpenSource Connections
 
Haystack 2019 Lightning Talk - The Future of Quepid - Charlie Hull
OpenSource Connections
 
Haystack 2019 Lightning Talk - State of Apache Tika - Tim Allison
OpenSource Connections
 
Haystack 2019 Lightning Talk - Relevance on 17 million full text documents - ...
OpenSource Connections
 
Haystack 2019 Lightning Talk - Solr Cloud on Kubernetes - Manoj Bharadwaj
OpenSource Connections
 
Haystack 2019 Lightning Talk - Quaerite a Search relevance evaluation toolkit...
OpenSource Connections
 
Haystack 2019 - Search-based recommendations at Politico - Ryan Kohl
OpenSource Connections
 
Haystack 2019 - Search with Vectors - Simon Hughes
OpenSource Connections
 
Haystack 2019 - Natural Language Search with Knowledge Graphs - Trey Grainger
OpenSource Connections
 
Haystack 2019 - Search Logs + Machine Learning = Auto-Tagging Inventory - Joh...
OpenSource Connections
 
Haystack 2019 - Improving Search Relevance with Numeric Features in Elasticse...
OpenSource Connections
 
Haystack 2019 - Architectural considerations on search relevancy in the conte...
OpenSource Connections
 
Haystack 2019 - Custom Solr Query Parser Design Option, and Pros & Cons - Ber...
OpenSource Connections
 
Haystack 2019 - Establishing a relevance focused culture in a large organizat...
OpenSource Connections
 
Haystack 2019 - Solving for Satisfaction: Introduction to Click Models - Eliz...
OpenSource Connections
 
2019 Haystack - How The New York Times Tackles Relevance - Jeremiah Via
OpenSource Connections
 

Recently uploaded (20)

PPTX
Digital-Transformation-Roadmap-for-Companies.pptx
David Malick Dieng
 
PDF
NewMind AI Monthly Chronicles - July 2025
NewMind AI
 
PDF
cuic standard and advanced reporting.pdf
SimoneTitaniumTomase
 
PDF
Approach and Philosophy of On baking technology
30anthuong17
 
PPT
“AI and Expert System Decision Support & Business Intelligence Systems”
ZubinRadhakrishnan
 
PDF
7 ChatGPT Prompts to Help You Define Your Ideal Customer Profile.pdf
SOFTTECHHUB
 
PPTX
MYSQL Presentation for SQL database connectivity
Swati270511
 
PPTX
Understanding_Digital_Forensics_Presentation.pptx
ImranKhan423233
 
PDF
Reach Out and Touch Someone: Haptics and Empathic Computing
Mark Billinghurst
 
PPTX
VMware vSphere Foundation How to Sell Presentation-Ver1.4-2-14-2024.pptx
blackmambaettijean
 
PDF
Network Security Unit 5.pdf for BCA BBA.
Serpent6
 
PDF
Dropbox Q2 2025 Financial Results & Investor Presentation
Dropbox
 
PPT
Teaching material agriculture food technology
LiaRayya
 
PDF
Diabetes mellitus diagnosis method based random forest with bat algorithm
IAESIJAI
 
PDF
Peak of Data & AI Encore- AI for Metadata and Smarter Workflows
Safe Software
 
PDF
NewMind AI Weekly Chronicles - August'25 Week I
NewMind AI
 
PDF
Advanced methodologies resolving dimensionality complications for autism neur...
IAESIJAI
 
PPTX
Effective Security Operations Center (SOC) A Modern, Strategic, and Threat-In...
ReZa AdineH
 
PDF
CIFDAQ's Market Insight: SEC Turns Pro Crypto
CIFDAQ
 
PDF
Chapter 3 Spatial Domain Image Processing.pdf
Getnet Tigabie Askale -(GM)
 
Digital-Transformation-Roadmap-for-Companies.pptx
David Malick Dieng
 
NewMind AI Monthly Chronicles - July 2025
NewMind AI
 
cuic standard and advanced reporting.pdf
SimoneTitaniumTomase
 
Approach and Philosophy of On baking technology
30anthuong17
 
“AI and Expert System Decision Support & Business Intelligence Systems”
ZubinRadhakrishnan
 
7 ChatGPT Prompts to Help You Define Your Ideal Customer Profile.pdf
SOFTTECHHUB
 
MYSQL Presentation for SQL database connectivity
Swati270511
 
Understanding_Digital_Forensics_Presentation.pptx
ImranKhan423233
 
Reach Out and Touch Someone: Haptics and Empathic Computing
Mark Billinghurst
 
VMware vSphere Foundation How to Sell Presentation-Ver1.4-2-14-2024.pptx
blackmambaettijean
 
Network Security Unit 5.pdf for BCA BBA.
Serpent6
 
Dropbox Q2 2025 Financial Results & Investor Presentation
Dropbox
 
Teaching material agriculture food technology
LiaRayya
 
Diabetes mellitus diagnosis method based random forest with bat algorithm
IAESIJAI
 
Peak of Data & AI Encore- AI for Metadata and Smarter Workflows
Safe Software
 
NewMind AI Weekly Chronicles - August'25 Week I
NewMind AI
 
Advanced methodologies resolving dimensionality complications for autism neur...
IAESIJAI
 
Effective Security Operations Center (SOC) A Modern, Strategic, and Threat-In...
ReZa AdineH
 
CIFDAQ's Market Insight: SEC Turns Pro Crypto
CIFDAQ
 
Chapter 3 Spatial Domain Image Processing.pdf
Getnet Tigabie Askale -(GM)
 

Encores

  • 1. Encores? - Going beyond matching and ranking of search results Berlin Buzzwords 2021 Eric Pugh, René Kriegler
  • 2. Who we are @renekrie @dep4b Combined 30 years of experience in search Open Source enthusiasts ASF member, Committers on: Solr, Querqy, SMUI, Quepid,
  • 4. Search - beyond matching and ranking We tend to focus on matching and ranking. Other search features are almost treated like an afterthought, like ‘encores’ that follow the main performance: ● Facets ● Query auto-completion ● Spelling correction ● Query relaxation => BUT: These are essential features that help the user formulate the query, understand and narrow down the results
  • 5. Our main act today (not encores!) ● Facets ● Query auto-completion ● Spelling correction ● Query relaxation Learn about solutions that come out of the box (in Solr) Typical challenges and how to overcome them Advanced solutions: understand the concepts, create your own
  • 6. The Art of Facets
  • 7. Facets help the user ... ● understand the search results (see ‘what is there’, learn about the domain) ● narrow down search results Chorus Electronics Project: https://github.com/querqy/chorus Try the Demo Ecommerce Shop: http://chorus.dev.o19s.com:4000/
  • 8. Facets help the user ... ● understand the search results (see ‘what is there’, learn about the domain) ● narrow down search results
  • 9. Challenges Getting the counts right in e-commerce search Showing the best facets in the best order Selecting the facet values to show
  • 10. “Qui numerare incipit errare incipit” Facet Counts
  • 11. Facets and filters A trivial example: query=t-shirts filter=color:black Challenge: We still need to count all colours in the facets, even if the search result contains only black t-shirts Solution: Tagging and exclusion of filters
  • 12. Facets and filters: tagging and exclusion Tagging: fq={!tag=f_color}color:black Exclusion: Facet param facet.field={!ex=f_color}color JSON facets "facet": { "color": { "type": "terms", "field": "color", "domain": { "excludeTags":"f_color" } } }
  • 13. Challenge: product variants Product ID: 9739, brand: “inteemate” Size: XS Price: 11.99 Size: XL Price: 11.99 Size: S Price: 12.99 Size: S Price: 12.99 Size: M Price: 13.99 Size: L Price: 13.99
  • 14. Challenge: product variants Product ID: 9739, brand: “inteemate” Size: XS Price: 11.99 Size: XL Price: 11.99 Size: S Price: 12.99 Size: S Price: 12.99 Size: M Price: 13.99 Size: L Price: 13.99 color: [green, yellow, blue] size: [XS, S, M, L, XL] price: [11.99, 12.99, 13.99] Merge into single document?? Facets would work great but false matches for filter color:green AND size:M
  • 15. Challenge: product variants Best solution (in our opinion): ● Index one document for each variant ● Group variants at query time using the collapse query parser: fq={!collapse field=productId} => Boolean filters work as expected => Great flexibility for counting facets => Fast enough
  • 16. Challenge: product variants Size: XS Price: 11.99 Product: 9739 Brand: inteemate Size: XL Price: 11.99 Product: 9739 Brand: inteemate Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: M Price: 13.99 Product: 9739 Brand: inteemate Size: L Price: 13.99 Product: 9739 Brand: inteemate filter query fq=brand:inteemate
  • 17. Challenge: product variants Size: XS Price: 11.99 Product: 9739 Brand: inteemate Size: XL Price: 11.99 Product: 9739 Brand: inteemate Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: M Price: 13.99 Product: 9739 Brand: inteemate Size: L Price: 13.99 Product: 9739 Brand: inteemate filter query fq=brand:inteemate filter query fq=color:blue
  • 18. Challenge: product variants Size: XS Price: 11.99 Product: 9739 Brand: inteemate Size: XL Price: 11.99 Product: 9739 Brand: inteemate Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: M Price: 13.99 Product: 9739 Brand: inteemate Size: L Price: 13.99 Product: 9739 Brand: inteemate filter query fq=brand:inteemate filter query fq=color:blue query q=t-shirt
  • 19. Challenge: product variants Size: XS Price: 11.99 Product: 9739 Brand: inteemate Size: XL Price: 11.99 Product: 9739 Brand: inteemate Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: M Price: 13.99 Product: 9739 Brand: inteemate Size: L Price: 13.99 Product: 9739 Brand: inteemate filter query fq=brand:inteemate filter query fq=color:blue query q=t-shirt post filter query fq={!collapse field=productId}
  • 20. Challenge: product variants in facets Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: M Price: 13.99 Product: 9739 Brand: inteemate Size: L Price: 13.99 Product: 9739 Brand: inteemate ... post filter query fq={!collapse field=productId} Facet counts will be correct for product attributes (“brand”)
  • 21. Challenge: product variants in facets Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: M Price: 13.99 Product: 9739 Brand: inteemate Size: L Price: 13.99 Product: 9739 Brand: inteemate ... post filter query fq={!collapse field=productId tag=coll} For facet counts of variant attributes we’ll have to tag and exclude collapse filter: "facet": { "size": { "type": "terms", "field": "size", "domain": { "excludeTags":"coll" } } } Sizes S, M, L all shown as ‘1 result’ in facets ✅
  • 22. Challenge: product variants in facets Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: M Price: 13.99 Product: 9739 Brand: inteemate Size: L Price: 13.99 Product: 9739 Brand: inteemate ... post filter query fq={!collapse field=productId tag=coll} For facet counts of variant attributes we’ll have to tag and exclude collapse filter: "facet": { "color": { "type": "terms", "field": "color", "domain": { "excludeTags":"coll" } } } Sizes S, M, L all shown as ‘1 result’ in facets ✅ Color ‘Blue’ shown as ‘3 results’ in facets ❌
  • 23. Challenge: product variants in facets Size: S Price: 12.99 Product: 9739 Brand: inteemate Size: M Price: 13.99 Product: 9739 Brand: inteemate Size: L Price: 13.99 Product: 9739 Brand: inteemate ... post filter query fq={!collapse field=productId tag=coll} For facet counts of variant attributes we’ll have to tag and exclude collapse filter: "facet": { "color": { "type": "terms", "field": "color", "domain": { "excludeTags":"coll" }, facet: { "numProducts":"unique(productId)" } } } Sizes S, M, L all shown as ‘1 result’ in facets ✅ Color ‘Blue’ shown as ‘1 result’ in facets ✅
  • 24. Collapse query parser - notes on implementation ● Beware of high cardinality of product IDs. ○ If you have 10M different product IDs in your index, the collapse query parser will allocate heap space for 2 arrays (float/int) x 10M elements (ca. 80 MB) per request! ○ Solution: ■ Many products have just 1 variant. It’s better to leave the productId empty in this case. ■ Combine with nullPolicy=expand, which avoids reserving array space for products without a productId: fq={!collapse field=productId nullPolicy=expand} ● All variants of a product must be indexed to the same shard
  • 26. Which facets should we show? Some domains are rich in attributes. For example, electronics could use 10k different attributes. Even if we reduced the number of attributes to be used in facets at index time, we could be left with several hundreds of candidates for facetting. Building a request for hundreds of facets is not feasible. We’ll show a simple solution, that will just use the search engine to select facets. At the other end of the spectrum, you could train a model, that predicts which facets to show for a given query.
  • 27. Which facets should we show? - Solution Index a field multivalued field that holds the names of the facettable fields... Doc1: screenSize: 17, .... facettableFields: [ “screenSize”, “ramGB”, “height”, “width”, ... ] Doc2: ... facettableFields: [ “screenSize”, “numHDMIPorts”, “height”, “width”, ... ]
  • 28. Which facets should we show? - Solution ... and execute an additional, prior facet request on this field. Add the facet values returned by this request as facet parameters to the main request: "facet": { "facettable_fields": { "type": "terms", "field": "facettable_fields" } } (query/filter queries are the same like in the ‘main request’) "facets":{ "facettable_fields":{ "buckets":[{ "val":"screenSize", "count":12}, { "val":"ramGB", "count":4}, { "val":"height", "count":3}, ] } }
  • 29. Which facets should we show? - Solution Index additional information together with the names of facettable fields facettableFields: [ "00010;screenSize;Screen size ", "00100;ramGB;Memory (GB) ", "00005;height;Height ", "00005;width;Width ", ...] Importance (padding makes values sortable!) Field name Label
  • 31. Which facet values? Category Pills being dynamically included IF the entropy model says they are meaningful for filtering the data
  • 34. Autocompletion - Using a Suggester <searchComponent name="suggest" class="solr.SuggestComponent"> <lst name="suggester"> <str name="name">mySuggester</str> <str name="lookupImpl">FuzzyLookupFactory</str> <str name="suggestAnalyzerFieldType">text_general</str> <str name="buildOnCommit">true</str> <str name="field">dictionary</str> </lst> </searchComponent> Experiment with combinations of Lookups & Dictionary implementations.
  • 35. Spellchecking - Using Solr component Two flavours: “cofffee --> coffee”, collations: “expresso machine” -->“espresso machine” <str name="spellcheck">true</str> <str name="spellcheck.dictionary">title</str> <str name="spellcheck.onlyMorePopular">true</str> <str name="spellcheck.extendedResults">true</str> <str name="spellcheck.collate">true</str> <str name="spellcheck.maxCollations">100</str> <str name="spellcheck.maxCollationTries">5</str> <str name="spellcheck.count">5</str> <str name="spellcheck.collateParam.mm">100%</str> <searchComponent name="suggest" class="solr.SuggestComponent"> <lst name="suggester"> <str name="name">mySuggester</str> <str name="lookupImpl">FuzzyLookupFactory</str> <str name="suggestAnalyzerFieldType">text_general</str> <str name="buildOnCommit">true</str> <str name="field">dictionary</str> </lst> </searchComponent> Good collations are what we want!
  • 36. Autocompletion - using a query index User enters: ja Show the best query completions in the best order! Using ideas from Joshua Bacher/Christine Bellstedt, Search Suggestions - The Underestimated Killer Feature of your Online Shop. Berlin Buzzwords 2018
  • 37. Autocompletion - using a query index User enters: ja Match prefix in field “match” q=*:*&fq=match:ja indexed with EdgeNGramFilter, lowercase, remove accents/ASCII folding, ... Optionally index and match spelling variants (jacket/jakcet)
  • 38. Autocompletion - using a query index Sort by “weight desc” q=*:*&fq=match:ja &sort=weight desc Sorting might get slow for short prefixes if the query index is large - tag the top N queries for lengths 1 and 2 and add another filter (fewer matches to sort, nicely cacheable): q=*:*&fq=match:ja &sort=weight desc &fq=top_len_2:true
  • 39. Autocompletion - using a query index If two queries have the same fingerprint, drop the one with the lower weight Fingerprint: concatenated sorted, normalised query tokens This increases the diversity of the suggestions.
  • 40. Autocompletion - using a query index Suggest the Labels as query completions!
  • 41. Autocompletion - using a query index You can show the best matching category for disambiguation and affirmation: * jacket * jacket in Fashion
  • 42. Spelling correction - using a query index Structure similar to query index for autocompletion Copy of the ‘match’ field indexed as n-grams
  • 43. Spelling correction - using a query index Filters on edit distance and rank based on n-grams (via TF*IDF) q=jakc jones &defType=edismax &qf=match_ngram &sow=false &fq=match:jakc jones~2
  • 44. Spelling correction - using a query index Add boost by weight (or a function of it) q=jakc jones &defType=edismax &qf=match_ngram &sow=false &fq=match:jakc jones~2 &boost=weight
  • 45. General model for spelling correction & autocompletion Noisy Channel Model / Bayesian Inference (Kernighan et. al., 1990; Jurafsky & Martin, 2009) Our ‘Weight’ field Edit distance, n-gram model, keyboard layout (Symspell!), ... prefix match for autocompletion
  • 47. Query relaxation Which query term should we drop if we can’t match all of them together? jacket xs green jacket xs green jacket xs green jacket xs green iphone 12 iphone 12 iphone 12
  • 48. Query relaxation - ‘mm’ anti-pattern Loosening ‘minimum should match’ (mm) constraint to < 100% iphone 12 iphone 12 You’ll get matches for “12” She will just see probably imprecise results that don’t match her query exactly. You cannot tell the user what happened and which term you dropped. She wouldn’t know what to do in order to improve the query. Don’t do this! At least not in e-commerce search
  • 49. Query relaxation - Solutions René.Kriegler, Query Relaxation - a rewriting technique between search and recommendations. Haystack Conference 2019
  • 50. Query relaxation - Solutions Try searching with each term individually and drop the one from the query that yields the fewest results (might require additional rules to avoid just keeping number terms)
  • 51. Query relaxation - Solutions Multi-layer Neural Network, Word embeddings as input to represent terms
  • 52. Encores? Facets, autocompletion, spelling correction, query relaxation are important features of a search application. We’ve shown simple out-of-the-box solutions and a path to implement more advanced approaches.