jhljx/RecSys-Industrial-Book

What it solves

This project addresses the need for a deep, systematic understanding of how industrial recommendation systems are architected and deployed, rather than just isolated algorithm tricks. It provides a multi-chapter book (PDFs) covering everything from system architecture, evaluation, retrieval, ranking, reranking, and blending modules, down to feature engineering, offline/online services, and bias challenges. It also tackles the practical problem of making such a dense technical resource navigable and actionable by pairing the book with an AI agent that can retrieve chapters, run structured courses, and verify learning progress.

How it works

The project is primarily a collection of PDF chapters (currently about 650 pages across upper and lower volumes, with roughly half complete). To make the book usable as a knowledge base, the README describes a companion AI agent package. This agent works entirely locally: you index the published PDFs (which are the authoritative text), and it builds a searchable index on your machine without exposing LaTeX or equivalent Markdown. Through command-line tools like recsys_agent.py, you can:

• Search across the 19 released chapters and get citations with exact page numbers.
• Start a structured 12-week learning path (new-engineer or graduate) that adapts to your feedback and available time.
• Submit evidence (actual command outputs, test scores, path reasoning) to prove mastery, not just claim it.
• Get adaptive review prompts, dependency graphs of topics, and self-evaluation.
• Use engineering workbenches to generate design contracts, review code, and debug recommendation pipelines.
• Run four hands-on template projects covering recall fusion, two-tower ANN, feature timing consistency, and overload/degradation handling.

All index and personal progress data stay on your local machine (ignored by Git). The agent also installs standard Agent Skills into compatible clients like Claude Code via tools/install_skills.py. The repository strictly separates the published PDFs from independently written chapter maps, concept cards, prerequisite graphs, and evidence-based learning paths—these are updated in lockstep with new chapters through an automated publishing workflow.

Who it’s for

The book is aimed at readers with a solid foundation in machine learning and recommendation basics: graduate students, algorithm engineers, architecture engineers, and technical researchers working on or studying industrial recommendation systems. It explicitly says it is not suitable for absolute beginners—prior knowledge of ML and deep learning is expected. The AI agent and templates are useful for anyone who wants to move from reading to practical, evidence-driven learning or who needs to build and debug real recommendation components.

Highlights

• A comprehensive, continuously updated open-source book covering industrial recommendation systems end-to-end (currently ~650 pages, about 50% of the full plan).
• Local, privacy-preserving AI agent: search, cite, and learn from the PDFs without any cloud dependency or MCP setup.
• Evidence-based learning: 12-week courses for new engineers and graduate-level researchers, with submission-based verification and adaptive review.
• Four runnable template projects that simulate core industrial scenarios like multi-channel recall, two-tower ANN, feature timing consistency, and overload handling.
• A publishing workflow that guarantees chapter updates come with corresponding knowledge maps, concept cards, and course unit updates, keeping all materials in sync.
• Compatible with Claude Code and other Agent Skills clients via a simple install script.