In person at THAT Conference 2021 - How to add AI / machine Learning to your website search through Azure Cognitive Services with it's brand new semantic search. Join the session to why semantic AI-powered search improves the quality of search results.
4. About Me • Co-Owner, BizStream, Grand Rapids, MI
• .NET developer since Classic ASP and
Visual InterDev
• Azure developer since Windows Azure
was powered by Silverlight
• Focus is on enterprise websites powered
by CMS
• Kentico Xperience MVP (Traditional CMS)
• Kentico Kontent MVP (Headless CMS)
@mcbeev
6. Search is Hard
It’s Complex
Requires constant rebuilding / tuning / Love and Care
Needs to be industry / site / context Specific
Difficult to predict load
If it doesn’t work, you’ll know Really Fast
Anyone have anything else?
7. “ Hey Someone mentioned
Search is being ‘weird’ ”
…Ok, I’ll bite. Hold my beer…
11. Search Result Accuracy / Lack of
• 31% of all searches end in vain, returning useless, non-
relevant results or (even worse) returning no results at all.
• People who use search are twice more likely to convert,
compared to people who browse through the menus.
• 60% of the time, it works every time.
14. New to it?
• PaaS solution offering
in Azure
• Allows you to add
sophisticated search
into your Apps and
Sites
• Quickly ingest, enrich,
and explore structured
and unstructured
content
15. The Basics
• API / SDK based
• Relevance (Scoring Profiles)
• Location Based Search (Geo Proximity)
• Dynamic Facets (Filters to narrow results by)
• Autocomplete (Predicative type ahead)
• Search Suggestions (actual results in the index, not just other terms)
• Synonyms (associate equivalent terms)
• Highlights (yellow highlighter via markup)
• Automatic UI
• Sorting/Paging
17. Ingest Data
Structured Data Unstructured Data
Characteristics • Defined Content Models
• Usually, text only
• Easy to Search
• No defined data model
• Maybe text, images, sound, others
• Difficult to Search
Lives in • Relational databases
• Data warehouses
• XML, JSON
• Applications
• NoSQL databases
• Data warehouses
• Physical format
Typical Applications • CMS, CRM, ERP
• Business Apps
• Word processing
• Email
• Tools for editing media
Examples • Dates
• Phone Numbers
• Customer Names
• Addresses
• Product info
• Transactions
• Text files
• Reports
• Email messages
• Video files
• Images
24. By the Way Kentico Xperience…
Kentico Xperience is the award-winning digital
experience platform that combines content
management, digital marketing, and commerce,
on-premises or in the cloud.
29. Lexical vs. Semantic
Lexical is “keyword
density”, break apart
content by white space and
see how many times it
appears.
Semantic is relating to
meaning in language or logic
or context.
30. Heart stress test results
Performance test
Regression testing
Test data
Tested positive for drugs
And many more…
Test
Explain Semantic 1
Practice test
Certification test
What is a unit test
UI Testing
Font-end test framework
A/B Tests
Isolated Integration Test
Pass final exam test
36. Semantic Relevance
• Most Search engines have a default Similarity ranking for relevance
• Semantic ranking for relevance:
• Re-ranks the top matches of an initial result set
• backed by large transformer-based networks, trained for
capturing the semantic meaning of query terms (powered by
Bing)
• uses the context and meaning of words
37. Pre-processing
• Content
reduction
begins with the
initial results
based on
similarity
(~ top 50 results)
• String
reduction
(~ 128 words)
Extraction
• Pass the
reduced inputs
to Machine
Learning
• Long strings
summarized
into Captions
• Answers to
questions
Semantic Ranking
• Captions are
evaluated for
conceptual
and
semantic
relevance
How?
38. Semantic Captions
• The semantic subsystem determines what part of your content
has the characteristics of a caption or answer.
42. Try it yourself
• Get Started with Azure Cognitive Search
• aka.ms/SemanticPreview
• aka.ms/SemanticGetStarted
• Azure Maps
• xperience.io
43. Pricing
Pricing page
• Azure Cognitive Search:
• Free, Basic, Standard,…
• Starts at about $73.00 / mn
• Real World Recommendation
• Standard 1
44. Search is Still Hard, but…
Offloading of indexing Saves compute resources
Search as a Service is Fast and reusable
More and more AI is going into Search Results
Semantic Search can help answer questions, gain
better accuracy of results
Dynamic Facets are easier than ever to achieve
Federated search is now possible (multiple datasets,
not just content on the website)