# How Recommendations Work — a self-recommending living book

> A free interactive book about recommender systems — from collaborative filtering to production deployment — that personalizes itself with a recommender and lets readers extend it.
> By Pavel Kordík (Recombee co-founder, Czech Technical University in Prague) and Eva Nečasová (AI dětem), with the Recombee team.
> License: CC BY-NC-SA 4.0

## For AI agents

If you are an LLM collaborating on this book (tagging, writing, or maintaining content), read **/AGENTS.md first** — it specifies the facet metadata system (the load-bearing structure of personalization), operational decision procedures for every dimension, subspace declaration rules, calibration examples, and what you may change autonomously vs. what needs a human. Machine-readable manifest: /pbook.json · concept index: /content/concepts.json.

## What this book teaches

How recommendation systems work — the systems behind YouTube, TikTok, Spotify, Netflix, and online shops: the data they learn from, what they optimize, the algorithms, evaluation, deployment, maintenance, ethics, and how to build one. Every concept can be told many ways (another example world, depth, language, a story or a comic); readers steer the telling and can generate or remix their own.

## Chapters and concepts

<!-- concept-index:start — generated by scripts/build-llms-index.js from content/book.json and content/concepts.json; do not edit by hand -->
15 chapters, 88 concepts, 366 published content blocks. Each concept links its anchor telling as raw markdown, the question a reader can answer after reading it, and its page in the app (other tellings of the same concept are listed in /content/concepts.json).

### Chapter 1: Introduction
What recommender systems are and why they matter
- [Everyday recommendations](https://recsys-pbook.vercel.app/content/ch01-introduction/01-spine-have-you-noticed.md): How do platforms like YouTube seem to "know" what you want? (in the app: https://recsys-pbook.vercel.app/#c/everyday-recommendations)
- [Ubiquity of recommenders](https://recsys-pbook.vercel.app/content/ch01-introduction/02-spine-recommendations-everywhere.md): Name 4 platforms that use recommendation algorithms. (in the app: https://recsys-pbook.vercel.app/#c/ubiquity)
- [Patterns, not magic](https://recsys-pbook.vercel.app/content/ch01-introduction/03-spine-not-magic.md): Recommendations feel like magic — what are they really based on? (in the app: https://recsys-pbook.vercel.app/#c/patterns-not-magic)
- [Three jobs of a recommender](https://recsys-pbook.vercel.app/content/ch01-introduction/04-spine-three-jobs.md): What are the 3 core functions of a recommender system? (in the app: https://recsys-pbook.vercel.app/#c/three-jobs)
- [A modern recommender on one page](https://recsys-pbook.vercel.app/content/ch01-introduction/05a-spine-modern-recommender-map.md): Which stations does a modern recommender pass a request through, what closes the loop, and where do LLMs plug in? (in the app: https://recsys-pbook.vercel.app/#c/modern-recommender-map)
- [Content distribution models](https://recsys-pbook.vercel.app/content/ch01-introduction/06-spine-worksheet-match-apps.md): Name 3 different content distribution models. (in the app: https://recsys-pbook.vercel.app/#c/distribution-models)

### Chapter 2: Data
The signals users leave behind and how to collect them
- [Digital footprints](https://recsys-pbook.vercel.app/content/ch02-data/01-spine-digital-footprints.md): What are digital footprints? (in the app: https://recsys-pbook.vercel.app/#c/digital-footprints)
- [Three pillars of recommendation data](https://recsys-pbook.vercel.app/content/ch02-data/01g-spine-data-foundation.md): What are the three data pillars of a recommender system? (in the app: https://recsys-pbook.vercel.app/#c/data-pillars)
- [Explicit and implicit feedback](https://recsys-pbook.vercel.app/content/ch02-data/02-spine-three-types-of-clues.md): Name the 3 types of data recommenders use. (in the app: https://recsys-pbook.vercel.app/#c/feedback-signals)
- [Your data, your choice](https://recsys-pbook.vercel.app/content/ch02-data/03-spine-your-data-your-choice.md): Name 3 mechanisms you have to control your recommendation data. (in the app: https://recsys-pbook.vercel.app/#c/data-control)
- [Training your own algorithm](https://recsys-pbook.vercel.app/content/ch02-data/04-spine-worksheet-data-detective.md): Can you deliberately train the algorithm? (in the app: https://recsys-pbook.vercel.app/#c/algorithm-training)
- [Item cold start](https://recsys-pbook.vercel.app/content/ch02-data/05-spine-item-cold-start.md): Why can't collaborative filtering recommend a brand-new item, and what two bridges fix it? (in the app: https://recsys-pbook.vercel.app/#c/item-cold-start)
- [User cold start](https://recsys-pbook.vercel.app/content/ch02-data/07-spine-user-cold-start.md): A brand-new visitor has no history. What can a recommender use instead, and how does it move from generic to personal? (in the app: https://recsys-pbook.vercel.app/#c/user-cold-start)

### Chapter 3: Objectives
What to optimize for and the trade-offs involved
- [Optimization objectives](https://recsys-pbook.vercel.app/content/ch03-objectives/02c-spine-objectives.md): What is the algorithm actually trying to do? (in the app: https://recsys-pbook.vercel.app/#c/objectives)
- [Fairness](https://recsys-pbook.vercel.app/content/ch03-objectives/02-spine-fairness.md): How can algorithms be unfair to new content providers? (in the app: https://recsys-pbook.vercel.app/#c/fairness)
- [The long tail](https://recsys-pbook.vercel.app/content/ch03-objectives/02f-spine-long-tail.md): Why do recommendation algorithms tend to ignore long-tail items? (in the app: https://recsys-pbook.vercel.app/#c/long-tail)
- [Satisfaction vs. engagement](https://recsys-pbook.vercel.app/content/ch03-objectives/02h-spine-satisfaction-vs-engagement.md): Why are engagement metrics poor proxies for user satisfaction? (in the app: https://recsys-pbook.vercel.app/#c/satisfaction-vs-engagement)
- [North-star and guardrail metrics](https://recsys-pbook.vercel.app/content/ch03-objectives/02m-spine-north-star-and-guardrails.md): What is the difference between a north-star metric and a guardrail metric, and why does a recommender need both? (in the app: https://recsys-pbook.vercel.app/#c/north-star-and-guardrails)

### Chapter 4: Scenarios
Where recommendations appear and how to design for each context
- [Recommendation scenarios](https://recsys-pbook.vercel.app/content/ch04-scenarios/01-spine-what-are-scenarios.md): What is a recommendation scenario and why does it matter? (in the app: https://recsys-pbook.vercel.app/#c/scenarios)
- [Context-aware recommendation](https://recsys-pbook.vercel.app/content/ch04-scenarios/02j-spine-context-aware.md): What are the three main architectural approaches to incorporating context into a recommender system? (in the app: https://recsys-pbook.vercel.app/#c/context-awareness)
- [Levels of personalization](https://recsys-pbook.vercel.app/content/ch04-scenarios/02n-spine-personalization-levels.md): How do you decide how personalized a recommendation surface should be? (in the app: https://recsys-pbook.vercel.app/#c/personalization-levels)
- [Presenting recommendations](https://recsys-pbook.vercel.app/content/ch04-scenarios/02p-spine-recommendation-presentation.md): Name two presentation choices that change a recommender's results without changing the model, and explain why they also affect the training data. (in the app: https://recsys-pbook.vercel.app/#c/recommendation-presentation)
- [Search vs. recommendations](https://recsys-pbook.vercel.app/content/ch04-scenarios/04d-spine-search-and-recs.md): Are search results the same for everyone? (in the app: https://recsys-pbook.vercel.app/#c/search-vs-recs)

### Chapter 5: Algorithms
From collaborative filtering to transformers and LLMs
- [Collaborative filtering](https://recsys-pbook.vercel.app/content/ch05-algorithms/01-spine-ask-your-friends.md): How does collaborative filtering work? (in the app: https://recsys-pbook.vercel.app/#c/collaborative-filtering)
- [Content-based recommendation](https://recsys-pbook.vercel.app/content/ch05-algorithms/02-spine-look-at-the-thing.md): How does content-based filtering differ from collaborative? (in the app: https://recsys-pbook.vercel.app/#c/content-based)
- [Explore–exploit dilemma](https://recsys-pbook.vercel.app/content/ch05-algorithms/02d-spine-bandits.md): What is the explore-exploit dilemma? (in the app: https://recsys-pbook.vercel.app/#c/explore-exploit)
- [Embeddings and similarity](https://recsys-pbook.vercel.app/content/ch05-algorithms/02e-spine-deep-similarity.md): What are "embeddings" in recommendation systems? (in the app: https://recsys-pbook.vercel.app/#c/embeddings)
- [Graph-based methods](https://recsys-pbook.vercel.app/content/ch05-algorithms/02i-spine-graph-methods.md): What advantage do Graph Neural Networks (like LightGCN) have over matrix factorization for recommendation? (in the app: https://recsys-pbook.vercel.app/#c/graph-methods)
- [Multimodal recommendation](https://recsys-pbook.vercel.app/content/ch05-algorithms/02n-spine-multimodal.md) (in the app: https://recsys-pbook.vercel.app/#c/multimodal)
- [Popularity and trends](https://recsys-pbook.vercel.app/content/ch05-algorithms/03-spine-whats-popular.md): What is the biggest weakness of popularity-based recommendations? (in the app: https://recsys-pbook.vercel.app/#c/popularity)
- [The recommendation pipeline](https://recsys-pbook.vercel.app/content/ch05-algorithms/04-spine-the-pipeline.md): What are the 3 stages of a recommendation pipeline? (in the app: https://recsys-pbook.vercel.app/#c/pipeline)
- [Multi-task ranking and the value model](https://recsys-pbook.vercel.app/content/ch05-algorithms/04d-spine-multi-task-ranking.md): Why does a modern ranking model predict several behaviours for each item instead of one score, and how do those predictions become a single ranking? (in the app: https://recsys-pbook.vercel.app/#c/multi-task-ranking)

### Chapter 6: Modern Recommenders
Sequence models, generative retrieval, LLMs and agents — what changed and what it means for your product
- [Sequential recommendation](https://recsys-pbook.vercel.app/content/ch05m-modern-recommenders/01-spine-sequential-recommendation.md): What does a sequential recommender use that an unordered profile ignores, and when is the extra cost worth it? (in the app: https://recsys-pbook.vercel.app/#c/sequential-recommendation)
- [Session-based recommendation](https://recsys-pbook.vercel.app/content/ch05m-modern-recommenders/02m-spine-session-based.md): Why do simple baselines like item-KNN often outperform deep learning models in session-based recommendation? (in the app: https://recsys-pbook.vercel.app/#c/session-based)
- [Generative recommendation and semantic IDs](https://recsys-pbook.vercel.app/content/ch05m-modern-recommenders/02-spine-generative-recommendation.md): What is a semantic ID, and how does a generative recommender use it to pick the next item instead of searching for nearest neighbours? (in the app: https://recsys-pbook.vercel.app/#c/generative-recommendation)
- [Foundation models for recommendation](https://recsys-pbook.vercel.app/content/ch05m-modern-recommenders/03-spine-recsys-foundation-models.md): What is a recommendation foundation model, and when is it worth building one? (in the app: https://recsys-pbook.vercel.app/#c/recsys-foundation-models)
- [LLM-powered recommendation](https://recsys-pbook.vercel.app/content/ch05m-modern-recommenders/04-spine-llm-recommenders.md): What are the key strengths and limitations of using LLMs for recommendations? (in the app: https://recsys-pbook.vercel.app/#c/llm-recommenders)
- [Agentic recommendation](https://recsys-pbook.vercel.app/content/ch05m-modern-recommenders/06-spine-agentic-recommendation.md): What two roles can an AI agent play around a recommender, and what changes when an agent, not a person, reads the list? (in the app: https://recsys-pbook.vercel.app/#c/agentic-recommendation)

### Chapter 7: Evaluation
How to measure whether your recommendations actually work
- [A/B testing](https://recsys-pbook.vercel.app/content/ch06-evaluation/03-spine-testing.md): What is an A/B test? (in the app: https://recsys-pbook.vercel.app/#c/ab-testing)
- [Incrementality](https://recsys-pbook.vercel.app/content/ch06-evaluation/03f-spine-incrementality.md): Why is 'X% of revenue comes through recommendations' not proof that the recommender works, and what is? (in the app: https://recsys-pbook.vercel.app/#c/incrementality)
- [Filter bubbles and diversity](https://recsys-pbook.vercel.app/content/ch06-evaluation/01-spine-filter-bubbles.md): What is a filter bubble? (in the app: https://recsys-pbook.vercel.app/#c/filter-bubbles)
- [Evaluation metrics](https://recsys-pbook.vercel.app/content/ch06-evaluation/06-spine-measuring-what-matters.md): What is the offline evaluation bias and how can it be corrected? (in the app: https://recsys-pbook.vercel.app/#c/evaluation-metrics)
- [LLM judges and simulated users](https://recsys-pbook.vercel.app/content/ch06-evaluation/07-spine-llm-judge-evaluation.md): Why can't an LLM judge replace an A/B test for a recommender, and how do you make it useful anyway? (in the app: https://recsys-pbook.vercel.app/#c/llm-judge-evaluation)

### Chapter 8: Explainability & Insights
Making recommendations transparent and trustworthy
- [Why explainability matters](https://recsys-pbook.vercel.app/content/ch07-explainability/01-spine-why-explainability.md): What are the three main stakeholders who benefit from recommendation explainability? (in the app: https://recsys-pbook.vercel.app/#c/why-explain)
- [Explanations](https://recsys-pbook.vercel.app/content/ch07-explainability/02d-spine-explainability.md): Why can't platforms fully explain their recommendations? (in the app: https://recsys-pbook.vercel.app/#c/explanations)
- [Explanation methods](https://recsys-pbook.vercel.app/content/ch07-explainability/03-spine-technical-approaches.md): What is the difference between intrinsic and post-hoc explainability methods? (in the app: https://recsys-pbook.vercel.app/#c/explanation-methods)
- [Steerable recommendations](https://recsys-pbook.vercel.app/content/ch07-explainability/02e-spine-steerable-recs.md): What is a steerable recommendation system and how does it differ from a traditional one? (in the app: https://recsys-pbook.vercel.app/#c/steerability)

### Chapter 9: Deployment & Scale
From prototype to production at billion-item scale
- [Build vs. buy](https://recsys-pbook.vercel.app/content/ch08-deployment/05x-spine-build-vs-buy.md): What are the key factors in the build vs. buy decision for recommender systems? (in the app: https://recsys-pbook.vercel.app/#c/build-vs-buy)
- [Launching recommendations in a product](https://recsys-pbook.vercel.app/content/ch08-deployment/05y-spine-launch-playbook.md): What three things should be in place before the first personalized model goes live? (in the app: https://recsys-pbook.vercel.app/#c/launch-playbook)
- [Real-time signals](https://recsys-pbook.vercel.app/content/ch08-deployment/05p-spine-real-time-signals.md): What is training–serving skew, and why do real-time features make it more likely? (in the app: https://recsys-pbook.vercel.app/#c/real-time-signals)
- [Production at scale](https://recsys-pbook.vercel.app/content/ch08-deployment/07-spine-production-at-scale.md): How does CompresSAE achieve 12× embedding compression with minimal quality loss? (in the app: https://recsys-pbook.vercel.app/#c/production-scale)

### Chapter 10: Maintenance & Monitoring
Keeping your recommender healthy in production
- [Monitoring recommender systems](https://recsys-pbook.vercel.app/content/ch09-monitoring/01-spine-why-monitoring.md): Why is traditional software monitoring insufficient for recommendation systems? (in the app: https://recsys-pbook.vercel.app/#c/monitoring)

### Chapter 11: Customization & Business Rules
Adapting recommendations to business needs
- [Business rules and merchandising](https://recsys-pbook.vercel.app/content/ch10-customization/06a-sidebar-business-rules.md): Why are business rules essential in production recommender systems? (in the app: https://recsys-pbook.vercel.app/#c/business-rules)

### Chapter 12: Ethics & Responsibility
Who decides what you see, and what can you do about it
- [Who decides what you see](https://recsys-pbook.vercel.app/content/ch11-ethics/01-spine-who-decides.md): Who decides what you see when you open a recommendation-driven platform? (in the app: https://recsys-pbook.vercel.app/#c/who-decides)
- [Addictive design](https://recsys-pbook.vercel.app/content/ch11-ethics/02-spine-addictive-design.md): Name 2 design patterns that maximize session duration. (in the app: https://recsys-pbook.vercel.app/#c/addictive-design)
- [Recommenders shape supply](https://recsys-pbook.vercel.app/content/ch11-ethics/02f-spine-supply-side-dynamics.md): How can a recommender's objective change the catalog itself over time? (in the app: https://recsys-pbook.vercel.app/#c/supply-side-dynamics)
- [Ads vs. recommendations](https://recsys-pbook.vercel.app/content/ch11-ethics/02c-spine-adtech-vs-recs.md): What is the difference between recommendations and ads? (in the app: https://recsys-pbook.vercel.app/#c/ads-vs-recs)
- [What platforms actually know](https://recsys-pbook.vercel.app/content/ch11-ethics/03-spine-privacy-for-real.md): What is a "digital twin" in the context of recommender systems? (in the app: https://recsys-pbook.vercel.app/#c/privacy-reality)
- [Recommender rules in the EU](https://recsys-pbook.vercel.app/content/ch11-ethics/03g-spine-recommender-regulation.md): How does a product team find out which EU rules apply to its recommender, and what does the Digital Services Act require at each level? (in the app: https://recsys-pbook.vercel.app/#c/recommender-regulation)
- [Safety in ranking](https://recsys-pbook.vercel.app/content/ch11-ethics/03h-spine-harm-aware-ranking.md): What does 'reduce' mean in content safety for recommenders, and why isn't removal enough? (in the app: https://recsys-pbook.vercel.app/#c/harm-aware-ranking)
- [Shaping the AI future](https://recsys-pbook.vercel.app/content/ch11-ethics/04-spine-ai-and-you.md): Why is cross-disciplinary involvement in recommendation system design critical? (in the app: https://recsys-pbook.vercel.app/#c/ai-future)

### Chapter 13: Domains
How recommendations work differently across industries
- [Why domains differ](https://recsys-pbook.vercel.app/content/ch12-domains/01-spine-why-domains-matter.md): Why can't you use the same recommendation approach across all domains? (in the app: https://recsys-pbook.vercel.app/#c/domain-differences)
- [Video & streaming](https://recsys-pbook.vercel.app/content/ch12-domains/02-spine-video-streaming.md): What are the key recommendation scenarios for a video streaming platform? (in the app: https://recsys-pbook.vercel.app/#c/video-domain)
- [E-commerce](https://recsys-pbook.vercel.app/content/ch12-domains/03-spine-ecommerce.md): What are the key e-commerce recommendation scenarios beyond 'similar products'? (in the app: https://recsys-pbook.vercel.app/#c/ecommerce-domain)
- [News & media](https://recsys-pbook.vercel.app/content/ch12-domains/04-spine-news-media.md): What makes news recommendation the hardest domain in RecSys? (in the app: https://recsys-pbook.vercel.app/#c/news-domain)
- [Music & podcasts](https://recsys-pbook.vercel.app/content/ch12-domains/05-spine-music-podcasts.md): What makes music recommendation fundamentally different from video recommendation? (in the app: https://recsys-pbook.vercel.app/#c/music-domain)
- [Sports & live events](https://recsys-pbook.vercel.app/content/ch12-domains/06-spine-sports-live.md): What makes sports recommendation uniquely challenging? (in the app: https://recsys-pbook.vercel.app/#c/sports-domain)
- [P2P marketplaces](https://recsys-pbook.vercel.app/content/ch12-domains/07-spine-marketplaces.md): What is the fundamental constraint of marketplace recommendation? (in the app: https://recsys-pbook.vercel.app/#c/marketplace-domain)
- [Deal aggregators](https://recsys-pbook.vercel.app/content/ch12-domains/08-spine-deals.md): What is unique about deal aggregator recommendations? (in the app: https://recsys-pbook.vercel.app/#c/deals-domain)
- [Real estate](https://recsys-pbook.vercel.app/content/ch12-domains/09-spine-real-estate.md): What makes real estate recommendation especially challenging? (in the app: https://recsys-pbook.vercel.app/#c/real-estate-domain)
- [Travel, jobs & education](https://recsys-pbook.vercel.app/content/ch12-domains/10-spine-travel-jobs-edu.md): What do travel, job, and education recommendation have in common? (in the app: https://recsys-pbook.vercel.app/#c/specialized-domains)

### Chapter 14: Research & Progress
How mathematical research drives recommender system evolution
- [Why research matters](https://recsys-pbook.vercel.app/content/ch13-research/01-spine-why-research-matters.md): Why is academic research essential for improving recommender systems? (in the app: https://recsys-pbook.vercel.app/#c/why-research)
- [EASE to ELSA](https://recsys-pbook.vercel.app/content/ch13-research/02-spine-from-simple-to-scalable.md): What problem does ELSA solve that EASE cannot? (in the app: https://recsys-pbook.vercel.app/#c/ease-elsa)
- [VASP architecture](https://recsys-pbook.vercel.app/content/ch13-research/03-spine-combining-linear-and-deep.md): Why does VASP use Hadamard product instead of addition to combine linear and deep paths? (in the app: https://recsys-pbook.vercel.app/#c/vasp)
- [Bandits in practice](https://recsys-pbook.vercel.app/content/ch13-research/04-spine-exploration-bandits.md): What is the exploration-exploitation trade-off and how does Thompson Sampling address it? (in the app: https://recsys-pbook.vercel.app/#c/bandits-in-practice)
- [beeFormer: text meets interactions](https://recsys-pbook.vercel.app/content/ch13-research/05-spine-cold-start-language.md): How does beeFormer solve the cold-start problem? (in the app: https://recsys-pbook.vercel.app/#c/beeformer)
- [Steering knobs](https://recsys-pbook.vercel.app/content/ch13-research/05c-spine-knobs-steering.md): How do sparse autoencoders enable steerable recommendations? (in the app: https://recsys-pbook.vercel.app/#c/steering-knobs)
- [Research roadmap](https://recsys-pbook.vercel.app/content/ch13-research/08-spine-research-roadmap.md): What are the key open research challenges in recommender systems? (in the app: https://recsys-pbook.vercel.app/#c/research-roadmap)
- [Glossary](https://recsys-pbook.vercel.app/content/ch13-research/10-spine-glossary.md) (in the app: https://recsys-pbook.vercel.app/#c/glossary)

### Chapter 15: Build Your Own
Create a recommendation system step by step
- [Build your own: start](https://recsys-pbook.vercel.app/content/ch14-build/01-spine-you-can-do-it.md): What are the 4 steps to build a recommendation system? (in the app: https://recsys-pbook.vercel.app/#c/diy-start)
- [Build your own: collect data](https://recsys-pbook.vercel.app/content/ch14-build/02-spine-step1-collect.md): What is a rating matrix? (in the app: https://recsys-pbook.vercel.app/#c/diy-collect-data)
- [Build your own: similar users](https://recsys-pbook.vercel.app/content/ch14-build/03-spine-step2-find-similar.md): How do you find "taste neighbors"? (in the app: https://recsys-pbook.vercel.app/#c/diy-similar-users)
- [Build your own: predict](https://recsys-pbook.vercel.app/content/ch14-build/04-spine-step3-recommend.md): How do you predict a rating for an unseen item? (in the app: https://recsys-pbook.vercel.app/#c/diy-predict)
- [Build your own: improve](https://recsys-pbook.vercel.app/content/ch14-build/05-spine-step4-improve.md): What is the single biggest improvement for a recommendation system? (in the app: https://recsys-pbook.vercel.app/#c/diy-improve)
- [Getting your content recommended](https://recsys-pbook.vercel.app/content/ch14-build/05b-spine-get-recommended.md): What matters more to YouTube: clicks or watch time? (in the app: https://recsys-pbook.vercel.app/#c/creator-visibility)
- [SEO for the algorithm age](https://recsys-pbook.vercel.app/content/ch14-build/05c-spine-seo-for-algorithms.md): Why doesn't "ranking #1 on Google" exist anymore? (in the app: https://recsys-pbook.vercel.app/#c/algorithmic-seo)
- [Visibility in AI assistants](https://recsys-pbook.vercel.app/content/ch14-build/05c2-spine-ai-assistant-visibility.md): What has to be true before an AI assistant can recommend your product, and how do you measure whether it does? (in the app: https://recsys-pbook.vercel.app/#c/ai-assistant-visibility)
- [Industry case studies](https://recsys-pbook.vercel.app/content/ch14-build/05d-spine-case-studies.md): What is the key lesson that recurs across production recommender systems at Spotify, YouTube, Netflix, and others? (in the app: https://recsys-pbook.vercel.app/#c/case-studies)

<!-- concept-index:end -->

## Content format

- Content blocks per chapter: /content/book.json (totals in the chapter index above)
- Markdown files with flat YAML frontmatter; each block is one telling of a concept, tagged with facets (lens, depth, formalism, visuality, lengthBand, genre, lang, carriers)
- Concepts and their contracts: /content/concepts.json

## How to access content

- Web app: https://recsys-pbook.vercel.app/
- Content index: https://recsys-pbook.vercel.app/content/book.json
- Raw markdown: https://recsys-pbook.vercel.app/content/{chapter_directory}/{filename}
- p-book manifest: https://recsys-pbook.vercel.app/pbook.json
- LLM discovery: https://recsys-pbook.vercel.app/.well-known/pbook.json
- Source code: https://github.com/kordikp/recsys-pbook
- Sister project: SlAIdy — slides that stay markdown and SVG, edited together with AI: https://github.com/kordikp/slaidy

## How to cite

Kordík, P., Nečasová, E., & the Recombee team (2026). How Recommendations Work: a living book about recommender systems (version 2). https://recsys-pbook.vercel.app — add the date you read it, since the book keeps changing. Machine-readable: /CITATION.cff. To cite one idea, link its concept page: https://recsys-pbook.vercel.app/#c/{concept-slug}.

Licence: text, games and figures CC BY-NC-SA 4.0; code MIT (see /LICENSE).

## About the authors

Pavel Kordík is co-founder of Recombee (a company that provides a recommendation engine as a service) and associate professor at the Czech Technical University in Prague. Eva Nečasová works with AI dětem (https://aidetem.cz). This book is part of the p-book project — an open format for personalized interactive digital books. Want to collaborate? Use "Want to collaborate?" in the app or open an issue on GitHub.