دانلود کتاب سیستم‌های RAG عامل‌محور (Agentic RAG) -طراحی، ساخت و استقرار عامل‌های هوش مصنوعیِ مبتنی بر دانش و آماده برای محیط عملیاتی؛ تسلط بر GraphRAG و جستجوی برداری (Vector Search)

This book exists because of a meeting that took place three months ago, in a windowless conference room at a Fortune 100 financial services company. Their AI engineering team had spent the better part of a year building a regulatory compliance RAG system over a corpus of 14 million internal documents. The system worked in demos. It impressed the board. It went to production. And then, three weeks after launch, it returned a hallucinated citation to a Securities and Exchange Commission rule that did not exist in the cited paragraph, in response to a query from a compliance officer preparing for a regulator meeting. The meeting I attended was the post-mortem. Around the table were six senior engineers who knew RAG well. They had read every major paper. They had implemented hybrid search, reranking, and query rewriting. They were using a top-tier vector database and a frontier-class language model. And yet, despite all of that, their system had failed in a way that — had it gone undetected — could have caused a regulatory incident. The failure was not at any single step. It was at the seams between steps. The chunker had broken the rule’s preamble away from its citation. The reranker, trained on a different corpus, had down-ranked the correct chunk. The query, written by the compliance officer in their own internal vocabulary, used terms that did not match the SEC’s phrasing. The language model, given a context window full of nearby-but-wrong passages, had constructed a confident answer that synthesised across documents in a way no rule actually authorised. No one chapter of any RAG book covered all of these failures at once. The papers covered individual techniques in isolation. The tutorials assumed naive pipelines. The vendor blog posts oversold their own components. What was missing was a production engineering manual that treated the entire retrieval pipeline as a system — one that could fail in any of dozens of ways, and that needed to be designed, monitored, evaluated, and continuously improved as a single coherent thing. That is what I have tried to write. The book is structured in six parts. Part I builds the foundations — what RAG actually is, why naive pipelines fail, and what the 2026 landscape looks like. Part II covers the core architecture: vector search, embedding engineering, chunking, indexing, hybrid retrieval, and reranking. Part III is the heart of the book — advanced patterns including GraphRAG, agentic RAG, self-correcting loops, and multi-agent retrieval orchestration. Part IV addresses production engineering: evaluation with RAGAS, retrieval observability, cost engineering, and the operational lifecycle. Part V is two deep enterprise case studies — financial compliance and healthcare/legal. Part VI is the 90-day roadmap for getting from where you are to production. Every chapter follows the same rhythm I developed in Multi-Agent AI Systems Handbook: a real production failure story to open, five subchapters that progress from problem to concept to implementation to trade-off, an anti-patterns section that names the specific mistakes I have seen destroy real systems, and a chapter summary that distils the takeaways. The voice is production-first. There is no hype. There are no promises that any single technique will save you. The promise of this book is more modest and more honest: if you build what is in these pages, with the discipline these patterns demand, you will ship retrieval systems that work — and that keep working when nobody is watching. A note on prerequisites. This book assumes comfort with Python, basic familiarity with language models and prompting, and some prior exposure to retrieval concepts at the level of having built or read about a naive RAG pipeline. You do not need a background in information retrieval theory. You do not need to be an ML researcher. You do need to be willing to think carefully about production failure modes — because that, more than any specific technique, is the discipline this book is trying to teach.