book

AI-Powered Search

Name: AI-Powered Search
Author: Trey Grainger
ISBN: 9781617296970

by Trey Grainger

January 2025

Intermediate to advanced

520 pages

16h 46m

English

Manning Publications

Audiobook available

Read now

Unlock full access

AI-Powered Search
copyright
contents
foreword
preface
acknowledgments
about this book
Who should read this bookHow this book is organized: A road mapAbout the codeliveBook discussion forumOther online resources
about the authors
about the cover illustration
Part 1 Modern search relevance

1 Introducing AI-powered search
1.1 What is AI-powered search?1.2 Understanding user intent1.2.1 What is a search engine?1.2.2 What do recommendation engines offer?1.2.3 The personalization spectrum between search and recommendations1.2.4 Semantic search and knowledge graphs1.2.5 Understanding the dimensions of user intent1.3 How does AI-powered search work?1.3.1 The core search foundation1.3.2 Reflected intelligence through feedback loops1.3.3 Signals boosting, collaborative filtering, and learning to rank1.3.4 Content and domain intelligence1.3.5 Generative AI and retrieval augmented generation1.3.6 Curated vs. black-box AI1.3.7 Architecture for an AI-powered search engineSummary
2 Working with natural language
2.1 The myth of unstructured data2.1.1 Types of unstructured data2.1.2 Data types in traditional structured databases2.1.3 Joins, fuzzy joins, and entity resolution in unstructured data2.2 The structure of natural language2.3 Distributional semantics and embeddings2.4 Modeling domain-specific knowledge2.5 Challenges in natural language understanding for search2.5.1 The challenge of ambiguity (polysemy)2.5.2 The challenge of understanding context2.5.3 The challenge of personalization2.5.4 Challenges interpreting queries vs. documents2.5.5 Challenges interpreting query intent2.6 Content + signals: The fuel powering AI-powered searchSummary
3 Ranking and content-based relevance
3.1 Scoring query and document vectors with cosine similarity3.1.1 Mapping text to vectors3.1.2 Calculating similarity between dense vector representations3.1.3 Calculating similarity between sparse vector representations3.1.4 Term frequency: Measuring how well documents match a term3.1.5 Inverse document frequency: Measuring the importance of a term in the query3.1.6 TF-IDF: A balanced weighting metric for text-based relevance3.2 Controlling the relevance calculation3.2.1 BM25: The industry standard default text-similarity algorithm3.2.2 Functions, functions, everywhere!3.2.3 Choosing multiplicative vs. additive boosting for relevance functions3.2.4 Differentiating matching (filtering) vs. ranking (scoring) of documents3.2.5 Logical matching: Weighting the relationships between terms in a query3.2.6 Separating concerns: Filtering vs. scoring3.3 Implementing user and domain-specific relevance rankingSummary
4 Crowdsourced relevance
4.1 Working with user signals4.1.1 Content vs. signals vs. models4.1.2 Setting up our product and signals datasets (RetroTech)4.1.3 Exploring the signals data4.1.4 Modeling users, sessions, and requests4.2 Introducing reflected intelligence4.2.1 What is reflected intelligence?4.2.2 Popularized relevance through signals boosting4.2.3 Personalized relevance through collaborative filtering4.2.4 Generalized relevance through learning to rank4.2.5 Other reflected intelligence models4.2.6 Crowdsourcing from contentSummary
Part 2 Learning domain-specific intent
5 Knowledge graph learning
5.1 Working with knowledge graphs5.2 Using our search engine as a knowledge graph5.3 Automatically extracting knowledge graphs from content5.3.1 Extracting arbitrary relationships from text5.3.2 Extracting hyponyms and hypernyms from text5.4 Learning intent by traversing semantic knowledge graphs5.4.1 What is a semantic knowledge graph?5.4.2 Indexing the datasets5.4.3 Structure of an SKG5.4.4 Calculating edge weights to measure the relatedness of nodes5.4.5 Using SKGs for query expansion5.4.6 Using SKGs for content-based recommendations5.4.7 Using SKGs to model arbitrary relationships5.5 Using knowledge graphs for semantic searchSummary
6 Using context to learn domain-specific language
6.1 Classifying query intent6.2 Query-sense disambiguation6.3 Learning related phrases from query signals6.3.1 Mining query logs for related queries6.3.2 Finding related queries through product interactions6.4 Phrase detection from user signals6.4.1 Treating queries as entities6.4.2 Extracting entities from more complex queries6.5 Misspellings and alternative representations6.5.1 Learning spelling corrections from documents6.5.2 Learning spelling corrections from user signals6.6 Pulling it all togetherSummary
7 Interpreting query intent through semantic search
7.1 The mechanics of query interpretation7.2 Indexing and searching on a local reviews dataset7.3 An end-to-end semantic search example7.4 Query interpretation pipelines7.4.1 Parsing a query for semantic search7.4.2 Enriching a query for semantic search7.4.3 Sparse lexical and expansion models7.4.4 Transforming a query for semantic search7.4.5 Searching with a semantically enhanced querySummary
Part 3 Reflected intelligence
8 Signals-boosting models
8.1 Basic signals boosting8.2 Normalizing signals8.3 Fighting signal spam8.3.1 Using signal spam to manipulate search results8.3.2 Combating signal spam through user-based filtering8.4 Combining multiple signal types8.5 Time decays and short-lived signals8.5.1 Handling time-insensitive signals8.5.2 Handling time-sensitive signals8.6 Index-time vs. query-time boosting: Balancing scale vs. flexibility8.6.1 Tradeoffs when using query-time boosting8.6.2 Implementing index-time signals boosting8.6.3 Tradeoffs when implementing index-time boostingSummary
9 Personalized search
9.1 Personalized search vs. recommendations9.1.1 Personalized queries9.1.2 User-guided recommendations9.2 Recommendation algorithm approaches9.2.1 Content-based recommenders9.2.2 Behavior-based recommenders9.2.3 Multimodal recommenders9.3 Implementing collaborative filtering9.3.1 Learning latent user and item features through matrix factorization9.3.2 Implementing collaborative filtering with Alternating Least Squares9.3.3 Personalizing search results with recommendation boosting9.4 Personalizing search using content-based embeddings9.4.1 Generating content-based latent features9.4.2 Implementing categorical guardrails for personalization9.4.3 Integrating embedding-based personalization into search results9.5 Challenges with personalizing search resultsSummary
10 Learning to rank for generalizable search relevance
10.1 What is LTR?10.1.1 Moving beyond manual relevance tuning10.1.2 Implementing LTR in the real world10.2 Step 1: A judgment list, starting with the training data10.3 Step 2: Feature logging and engineering10.3.1 Storing features in a modern search engine10.3.2 Logging features from our search engine corpus10.4 Step 3: Transforming LTR to a traditional machine learning problem10.4.1 SVMrank: Transforming ranking to binary classification10.4.2 Transforming our LTR training task to binary classification10.5 Step 4: Training (and testing!) the model10.5.1 Turning a separating hyperplane’s vector into a scoring function10.5.2 Taking the model for a test drive10.5.3 Validating the model10.6 Steps 5 and 6: Upload a model and search10.6.1 Deploying and using the LTR model10.6.2 A note on LTR performance10.7 Rinse and repeatSummary
11 Automating learning to rank with click models
11.1 (Re)creating judgment lists from signals11.1.1 Generating implicit, probabilistic judgments from signals11.1.2 Training an LTR model using probabilistic judgments11.1.3 Click-Through Rate: Your first click model11.1.4 Common biases in judgments11.2 Overcoming position bias11.2.1 Defining position bias11.2.2 Position bias in RetroTech data11.2.3 Simplified dynamic Bayesian network: A click model that overcomes position bias11.3 Handling confidence bias: Not upending your model due to a few lucky clicks11.3.1 The low-confidence problem in click data11.3.2 Using a beta prior to model confidence probabilistically11.4 Exploring your training data in an LTR systemSummary
12 Overcoming ranking bias through active learning
12.1 Our automated LTR engine in a few lines of code12.1.1 Turning clicks into training data (chapter 11 in one line of code)12.1.2 Model training and evaluation in a few function calls12.2 A/B testing a new model12.2.1 Taking a better model out for a test drive12.2.2 Defining an A/B test in the context of automated LTR12.2.3 Graduating the better model into an A/B test12.2.4 When “good” models go bad: What we can learn from a failed A/B test12.3 Overcoming presentation bias: Knowing when to explore vs. exploit12.3.1 Presentation bias in the RetroTech training data12.3.2 Beyond the ad hoc: Thoughtfully exploring with a Gaussian process12.3.3 Examining the outcome of our explorations12.4 Exploit, explore, gather, rinse, repeat: A robust automated LTR loopSummary
Part 4 The search frontier
13 Semantic search with dense vectors
13.1 Representation of meaning through embeddings13.2 Search using dense vectors13.2.1 A brief refresher on sparse vectors13.2.2 A conceptual dense vector search engine13.3 Getting text embeddings by using a Transformer encoder13.3.1 What is a Transformer?13.3.2 Openly available pretrained Transformer models13.4 Applying Transformers to search13.4.1 Using the Stack Exchange outdoors dataset13.4.2 Fine-tuning and the Semantic Text Similarity Benchmark13.4.3 Introducing the SBERT Transformer library13.5 Natural language autocomplete13.5.1 Getting noun and verb phrases for our nearest-neighbor vocabulary13.5.2 Getting embeddings13.5.3 ANN search13.5.4 ANN index implementation13.6 Semantic search with LLM embeddings13.6.1 Getting titles and their embeddings13.6.2 Creating and searching the nearest-neighbor index13.7 Quantization and representation learning for more efficient vector search13.7.1 Scalar quantization13.7.2 Binary quantization13.7.3 Product quantization13.7.4 Matryoshka Representation Learning13.7.5 Combining multiple vector search optimization approaches13.8 Cross-encoders vs. bi-encodersSummary
14 Question answering with a fine-tuned large language model
14.1 Question-answering overview14.1.1 How a question-answering model works14.1.2 The retriever-reader pattern14.2 Constructing a question-answering training dataset14.2.1 Gathering and cleaning a question-answering dataset14.2.2 Creating the silver set: Automatically labeling data from a pretrained model14.2.3 Human-in-the-loop training: Manually correcting the silver set to produce a golden set14.2.4 Formatting the golden set for training, testing, and validation14.3 Fine-tuning the question-answering model14.3.1 Tokenizing and shaping our labeled data14.3.2 Configuring the model trainer14.3.3 Performing training and evaluating loss14.3.4 Holdout validation and confirmation14.4 Building the reader with the new fine-tuned model14.5 Incorporating the retriever: Using the question-answering model with the search engine14.5.1 Step 1: Querying the retriever14.5.2 Step 2: Inferring answers from the reader model14.5.3 Step 3: Reranking the answers14.5.4 Step 4: Returning results by combining the retriever, reader, and rerankerSummary
15 Foundation models and emerging search paradigms
15.1 Understanding foundation models15.1.1 What qualifies as a foundation model?15.1.2 Training vs. fine-tuning vs. prompting15.2 Generative search15.2.1 Retrieval augmented generation15.2.2 Results summarization using foundation models15.2.3 Data generation using foundation models15.2.4 Evaluating generative output15.2.5 Constructing your own metric15.2.6 Algorithmic prompt optimization15.3 Multimodal search15.3.1 Common modes for multimodal search15.3.2 Implementing multimodal search15.4 Other emerging AI-powered search paradigms15.4.1 Conversational and contextual search15.4.2 Agent-based search15.5 Hybrid search15.5.1 Reciprocal rank fusion15.5.2 Other hybrid search algorithms15.6 Convergence of contextual technologies15.7 All the above, please!Summary
appendix A Running the code examples
A.1 Overall structure of code examplesA.2 Pulling the source codeA.3 Building and running the codeA.4 Working with JupyterA.5 Working with Docker
appendix B Supported search engines and vector databases
B.1 Supported enginesB.2 Swapping out the engineB.3 The engine and collection abstractionsB.4 Adding support for additional engines
index

Content preview from AI-Powered Search

12 Overcoming ranking bias through active learning

This chapter covers

Harnessing live user interactions to gather feedback on a deployed LTR model
A/B testing search relevance solutions with live users
Using active learning to explore potentially relevant results beyond the top results
Balancing exploiting user interactions while exploring what else might be relevant

So far, our learning to rank (LTR) work has taken place in the lab. In previous chapters, we built models using automatically constructed training data from user clicks. In this chapter, we’ll take our model into the real world for a test drive with (simulated) live users!

Recall that we compared an automated LTR system to a self-driving car. Internally, the car has an ...

Become an O’Reilly member and get unlimited access to this title plus top books and audiobooks from O’Reilly and nearly 200 top publishers, thousands of courses curated by job role, 150+ live events each month,
and much more.

Read now

Unlock full access

More than 5,000 organizations count on O’Reilly

O’Reilly covers everything we've got, with content to help us build a world-class technology community, upgrade the capabilities and competencies of our teams, and improve overall team performance as well as their engagement.

Julian F.

Head of Cybersecurity

I wanted to learn C and C++, but it didn't click for me until I picked up an O'Reilly book. When I went on the O’Reilly platform, I was astonished to find all the books there, plus live events and sandboxes so you could play around with the technology.

Addison B.

Field Engineer

I’ve been on the O’Reilly platform for more than eight years. I use a couple of learning platforms, but I'm on O'Reilly more than anybody else. When you're there, you start learning. I'm never disappointed.

Amir M.

Data Platform Tech Lead

I'm always learning. So when I got on to O'Reilly, I was like a kid in a candy store. There are playlists. There are answers. There's on-demand training. It's worth its weight in gold, in terms of what it allows me to do.

Mark W.

Embedded Software Engineer

Publisher Resources

ISBN: 9781617296970Publisher Support Publisher Website

Cloud Computing

Data Engineering

Data Science

AI & ML

Programming Languages

Software Architecture

IT/Ops

Security

Design

Business

Soft Skills

AI-Powered Search

by Trey Grainger