AI Agents and Applications With LangChain, LangGraph, and MCP (Roberto Infante)（Z-Library）

M A N N I N G Roberto Infante With LangChain, LangGraph and MCP

2 EPILOGUE User question Retrieval Prompt Generation User User question Vector DB Retriever User question Retrieved chunks Prompt LLM Response Chatbot 1. The retriever transforms the user question into the corresponding embedding. 2. The retriever searches for relevant chunks using the question embedding. 3. The vector DB returns text chunks relevant to the user question. 4. The chatbot creates a prompt with the user question and the retrieved chunks (the context). 5. The LLM accepts the prompt and synthesizes a corresponding completion (the response). Similar chunks Question embeddings Retrieval-Augmented Generation (RAG) Q&A stage: retrieval and generation RAG Q&A Stage Licensed to THIAGO BANDEIRA <thiago@lar.ifce.edu.br>

AI Agents and ApplicationsLicensed to THIAGO BANDEIRA <thiago@lar.ifce.edu.br>

iiLicensed to THIAGO BANDEIRA <thiago@lar.ifce.edu.br>

AI Agents and Applications WITH LANGCHAIN, LANGGRAPH AND MCP ROBERTO INFANTE M A N N I N G SHELTER ISLAND Licensed to THIAGO BANDEIRA <thiago@lar.ifce.edu.br>

For online information and ordering of this and other Manning books, please visit www.manning.com. The publisher offers discounts on this book when ordered in quantity. For more information, please contact Special Sales Department Manning Publications Co. 20 Baldwin Road PO Box 761 Shelter Island, NY 11964 Email: orders@manning.com ©2026 by Manning Publications Co. All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by means electronic, mechanical, photocopying, or otherwise, without prior written permission of the publisher. Many of the designations used by manufacturers and sellers to distinguish their products are claimed as trademarks. Where those designations appear in the book, and Manning Publications was aware of a trademark claim, the designations have been printed in initial caps or all caps. Recognizing the importance of preserving what has been written, it is Manning’s policy to have the books we publish printed on acid-free paper, and we exert our best efforts to that end. Recognizing also our responsibility to conserve the resources of our planet, Manning books are printed on paper that is at least 15 percent recycled and processed without the use of elemental chlorine. The author and publisher have made every effort to ensure that the information in this book was correct at press time. The author and publisher do not assume and hereby disclaim any liability to any party for any loss, damage, or disruption caused by errors or omissions, whether such errors or omissions result from negligence, accident, or any other cause, or from any usage of the information herein. Manning Publications Co. Development editor: Dustin Archibald 20 Baldwin Road Technical editors: Keerthivasan Santhanakrishnan PO Box 761 and Antowan Malik Batts Shelter Island, NY 11964 Review editor: Kishor Rit Production editor: Kathy Rossland Copy editor: Julie McNamee Proofreader: Melody Dolab Technical proofreader: Andrew R. Freed Typesetter and cover designer: Marija TudorISBN 9781633436541 Printed in the United States of America Licensed to THIAGO BANDEIRA <thiago@lar.ifce.edu.br>

To my mother and father Licensed to THIAGO BANDEIRA <thiago@lar.ifce.edu.br>

brief contents PART 1 GETTING STARTED WITH LLMS ....................................... 1 1 ■ Introduction to AI agents and applications 3 2 ■ Executing prompts programmatically 27 PART 2 SUMMARIZATION ........................................................... 53 3 ■ Summarizing text using LangChain 55 4 ■ Building a research summarization engine 69 5 ■ Agentic workflows with LangGraph 103 PART 3 Q&A CHATBOTS ......................................................... 119 6 ■ RAG fundamentals with ChromaDB 121 7 ■ Q&A chatbots with LangChain and LangSmith 143 PART 4 ADVANCED RAG ......................................................... 171 8 ■ Advanced indexing 173 9 ■ Question transformations 205 10 ■ Query generation, routing, and retrieval postprocessing 228 PART 5 AI AGENTS .................................................................. 265 11 ■ Building tool-based agents with LangGraph 267vi 12 ■ Multi-agent systems 293 Licensed to THIAGO BANDEIRA <thiago@lar.ifce.edu.br>

BRIEF CONTENTS vii 13 ■ Building and consuming MCP servers 308 14 ■ Productionizing AI agents: Memory, guardrails, and beyond 327 appendix A Trying out LangChain 351 appendix B Setting up a Jupyter Notebook environment 357 appendix C Choosing an LLM 360 appendix D Installing SQLite on Windows 375 appendix E Open source LLMs 377Licensed to THIAGO BANDEIRA <thiago@lar.ifce.edu.br>

contents preface xvi acknowledgments xviii about this book xx about the author xxv about the cover illustration xxvi PART 1 GETTING STARTED WITH LLMS ........................ 1 1 Introduction to AI agents and applications 3 1.1 Building LLM-based applications and agents 4 LLM-based applications: Summarization and Q&A engines 4 ■ LLM-based chatbots 7 ■ AI agents 8 1.2 Introducing LangChain 11 LangChain architecture 12 ■ LangChain’s core object model 15 1.3 Typical LLM use cases 19 1.4 How to adapt an LLM to your needs 20 Prompt engineering 20 ■ RAG 20 ■ Fine-tuning 22 1.5 Which LLMs to choose 23viii 1.6 What you’ll learn from this book 24 Licensed to THIAGO BANDEIRA <thiago@lar.ifce.edu.br>

CONTENTS ix 2 Executing prompts programmatically 27 2.1 Running prompts programmatically 28 Setting up the environment for this chapter 28 ■ Minimal prompt execution 30 2.2 Running prompts with LangChain 31 2.3 Prompt templates 32 Implementing a prompt template with a Python function 32 Using LangChain’s PromptTemplate 33 2.4 Prompt types 34 Text classification 34 ■ Sentiment analysis 35 ■ Text summarization 36 ■ Composing text 36 ■ Question answering 38 ■ Reasoning 39 2.5 Reasoning in detail 40 One-shot learning 40 ■ Two-shot learning 41 ■ Providing steps 41 ■ Few-shot learning 42 ■ Implementing few-shot learning with LangChain 44 ■ Chain of Thought 46 2.6 Prompt structure 48 PART 2 SUMMARIZATION ............................................ 53 3 Summarizing text using LangChain 55 3.1 Summarizing a document bigger than the context window 56 Chunking the text into Document objects 57 ■ Split 58 Map 59 ■ Reduce 60 ■ MapReduce combined chain 60 MapReduce execution 61 3.2 Summarizing across documents 61 Creating a list of Document objects 63 ■ Wikipedia content 63 File-based content 64 ■ Creating the Document list 65 Progressively refining the final summary 65 3.3 Summarization flowchart 67 4 Building a research summarization engine 69 4.1 Overview of a research summarization engine 70 4.2 Setting up the project 71 4.3 Implementing the core functionality 73 Implementing web searching 73 ■ Implementing web scraping 74 ■ Instantiating the LLM client 75 JSON to Python object converter 76 Licensed to THIAGO BANDEIRA <thiago@lar.ifce.edu.br>

CONTENTSx 4.4 Enhancing the architecture with query rewriting 76 4.5 Prompt engineering 78 Crafting web search prompts 78 ■ Crafting summarization prompts 81 ■ Research report prompt 81 4.6 Initial implementation 82 Importing functions and prompt templates 82 ■ Setting constants and input variables 83 ■ Instantiating the LLM client 83 ■ Generating the web searches and collecting the results 83 ■ Scraping the web results 85 ■ Summarizing the web results 86 ■ Generating the research report 87 4.7 Reimplementing the research summary engine in LCEL 89 Assistant Instructions chain 91 ■ Web Searches chain 93 Search and Summarization chain 94 ■ Web Research chain 99 5 Agentic workflows with LangGraph 103 5.1 Understanding agentic workflows and agents 104 Workflows 105 ■ Agents 106 ■ When to use agent-based architectures 106 ■ Agent development frameworks 106 5.2 LangGraph basics 107 5.3 Moving from LangChain chains to LangGraph 107 5.4 LangGraph core components 108 StateGraph structure 109 ■ State management and typing 109 Node functions and edge definitions 110 ■ Entry points and end conditions 110 5.5 Turning the web research assistant into an AI agent 110 Original LangChain implementation overview 111 ■ Identifying components for conversion 112 ■ Step-by-step transformation process 112 ■ Code comparison and benefits realized 116 PART 3 Q&A CHATBOTS.......................................... 119 6 RAG fundamentals with ChromaDB 121 6.1 Semantic search 122 A basic Q&A chatbot over a single document 122 ■ A more complex Q&A chatbot over a knowledge base 126 ■ The RAG design pattern 127 6.2 Vector stores 130 What’s a vector store? 130 ■ How do vector stores work? 131 Vector libraries vs. vector databases 131 ■ Most popular vector Licensed to THIAGO BANDEIRA <thiago@lar.ifce.edu.br>

CONTENTS xi stores 132 ■ Storing text and performing a semantic search using Chroma 133 6.3 Implementing RAG from scratch 136 Retrieving content from the vector database 137 ■ Invoking the LLM 137 ■ Building the chatbot 139 ■ Recap of RAG terminology 140 7 Q&A chatbots with LangChain and LangSmith 143 7.1 LangChain object model for Q&A chatbots 144 Content ingestion (indexing) stage 144 ■ Q&A (retrieval and generation) stage 146 7.2 Vector store content ingestion 148 Splitting and storing the documents 149 ■ Removing duplication 150 ■ Ingesting multiple documents from a folder 151 7.3 Q&A across stored documents 152 Querying the vector store directly 153 ■ Asking a question through a LangChain chain 153 ■ Completing the RAG chain setup 154 ■ Follow-up question 156 7.4 Chatbot memory of message history 157 Amending the prompt 158 ■ Updating the chat message history 159 ■ Feeding the chat history to the RAG chain 160 Putting everything together 160 7.5 Tracing execution with LangSmith 163 PART 4 ADVANCED RAG .......................................... 171 8 Advanced indexing 173 8.1 Improving RAG accuracy 174 Content ingestion stage 174 ■ Question-answering stage 175 8.2 Advanced document indexing 177 8.3 Splitting strategy 177 Splitting strategies 178 ■ Factors to consider 179 ■ Choosing the right strategy 179 ■ Splitting by HTML header 179 8.4 Embedding strategy 183 Embedding child chunks with ParentDocumentRetriever 184 Embedding child chunks with MultiVectorRetriever 187 Licensed to THIAGO BANDEIRA <thiago@lar.ifce.edu.br>

CONTENTSxii Embedding document summaries 189 ■ Embedding hypothetical questions 193 8.5 Granular chunk expansion 197 8.6 Semi-structured content 201 8.7 Multimodal RAG 202 9 Question transformations 205 9.1 Rewrite-Retrieve-Read 206 Retrieving content using the original user question 209 Setting up the query rewriter chain 210 ■ Retrieving content with the rewritten query 211 ■ Combining everything into a single RAG chain 212 9.2 Generating multiple queries 213 Setting up the chain for generating multiple queries 215 Setting up a custom multi-query retriever 216 ■ Using a standard MultiQueryRetriever instance 218 9.3 Step-back question 218 Setting up the chain to generate a step-back question 220 Incorporating step-back question generation into the RAG chain 220 9.4 Hypothetical Document Embeddings (HyDE) 222 Generating a hypothetical document for the user question 223 Integrating the HyDE chain into the RAG chain 223 9.5 Single-step and multi-step decomposition 224 10 Query generation, routing, and retrieval postprocessing 228 10.1 Content database query generation 229 10.2 Self-querying (metadata query enrichment) 230 Ingestion: Metadata enrichment 232 ■ Q&A on a metadata- enriched collection 234 10.3 Generating a structured SQL query 239 Installing SQLite 240 ■ Setting up and connecting to the database 240 ■ Generating SQL queries from natural language 242 ■ Executing the SQL query 244 10.4 Generating a semantic SQL query 244 Standard SQL query 245 ■ Semantic SQL query 246 Creating the embeddings 246 ■ Performing a semantic SQL search 247 ■ Automating semantic SQL search 247 Benefits of a semantic SQL search 247 Licensed to THIAGO BANDEIRA <thiago@lar.ifce.edu.br>

CONTENTS xiii 10.5 Generating queries for a graph database 248 10.6 Chain routing 251 Setting up data retrievers 252 ■ Setting up the query router 252 Integrating the chain router into a full RAG chain 254 10.7 Retrieval postprocessing 256 Similarity postprocessors 257 ■ Keyword postprocessors 257 Time weighting 257 ■ RAG fusion (Reciprocal Rank Fusion) 258 PART 5 AI AGENTS ................................................... 265 11 Building tool-based agents with LangGraph 267 11.1 Starting simple: Building a single-tool travel info agent 268 Project setup 268 ■ Loading environment variables 269 Preparing the travel information vector store 269 11.2 Enabling agents to call tools 271 From function calling to tool calling 271 ■ How tool calling works with LLMs 273 ■ Registering tools with the LLM 273 Agent state: Tracking the conversation 275 ■ Executing tool calls 275 ■ The LLM node: Coordinating reasoning and action 277 11.3 Assembling the agent graph 277 11.4 Understanding the agent graph structure 278 11.5 Running the agent chatbot: The Read-Eval-Print Loop 279 11.6 Executing a request 279 Step-by-step debugging 280 11.7 Expanding your agent: Adding a weather forecast tool 283 Implementing a mock weather service 283 ■ Creating the weather forecast tool 284 ■ Updating the agent for multi-tool support 284 11.8 Executing the multi-tool agent 284 Running the multi-tool agent (initial behavior) 285 ■ Improving LLM tool usage with system guidance 285 11.9 Using prebuilt components for rapid development 288 Refactoring to use the LangGraph ReAct agent 288 Running the prebuilt agent 289 ■ Observing and debugging with LangSmith 289 ■ Enabling LangSmith tracing 289Licensed to THIAGO BANDEIRA <thiago@lar.ifce.edu.br>

CONTENTSxiv 12 Multi-agent systems 293 12.1 Building an accommodation booking agent 294 Hotel booking tool 294 ■ B&B booking tool 295 ReAct accommodation booking agent 297 12.2 Building a router-based travel assistant 298 Designing the router agent 298 ■ Routing logic 298 Building the multi-agent graph 300 ■ Trying out the router agent 301 12.3 Handling multi-agent requests with a Supervisor component 302 The Supervisor pattern: An agent of agents 302 ■ From “one-way” to “return ticket” interactions 304 ■ Trying out the Supervisor agent 305 13 Building and consuming MCP servers 308 13.1 Introduction to MCP servers 309 The problem: Context integration at scale 309 ■ The solution: The Model Context Protocol 310 ■ The MCP ecosystem 311 13.2 How to build MCP servers 312 Essential resources for MCP server development 312 ■ Official language-specific MCP SDKs 312 ■ Consuming MCP servers in LLM applications and agents 313 13.3 Building a weather MCP server 314 Implementing the MCP server 314 ■ Trying out the MCP server with MCP Inspector 316 ■ Consuming the MCP server from a test MCP host 320 13.4 Integrating the Weather MCP tool into an agent 322 Preparing the travel agent for live weather data 323 Integrating the AccuWeather MCP tool 323 ■ Updating the agent chat loop 323 ■ Combining local and remote tools 324 Testing and verification 324 ■ Using the agent for complex queries 325 14 Productionizing AI agents: Memory, guardrails, and beyond 327 14.1 Memory 328 Types of memory 328 ■ Why short-term memory is needed 328 Checkpoints in LangGraph 329 ■ Adding short-term memory to Licensed to THIAGO BANDEIRA <thiago@lar.ifce.edu.br>

CONTENTS xv our travel assistant 331 ■ Executing the checkpointer-enabled assistant 334 ■ Rewinding the state to a past checkpoint 334 14.2 Guardrails 337 Implementing guardrails to reject nontravel-related questions 338 Implementing more restrictive guardrails at the agent level 343 14.3 Beyond this chapter 346 Long-term user and application memory 346 ■ Human-in-the- loop 346 ■ Post-model guardrails 347 ■ Evaluation of AI agents and applications 347 ■ Deployment on the LangGraph platform and Open Agent Platform 348 appendix A Trying out LangChain 351 appendix B Setting up a Jupyter Notebook environment 357 appendix C Choosing an LLM 360 appendix D Installing SQLite on Windows 375 appendix E Open source LLMs 377 index 411Licensed to THIAGO BANDEIRA <thiago@lar.ifce.edu.br>

preface In late 2022, something changed. Large language models (LLMs) stopped feeling like experimental demonstrations and started becoming genuinely useful. A quick attempt to summarize a paragraph evolved into a chatbot capable of answering ques- tions, and a small script turned into a service that other teams wanted to try. Before long, LLMs were no longer a curiosity—they had become an essential part of the soft- ware development toolkit. Here's why that’s so exciting: LLMs allow software to “speak human.” They can revise a contract, turn logs into meaningful answers, draft code, and plan the next step—and then invoke the right tools and data to actually accomplish the task. Com- bined with retrieval and tool use, an application stops feeling like a rigid machine and begins to feel more like a collaborative partner. The potential is significant, but turn- ing that potential into production systems isn’t simple. It still requires careful work to integrate data flows, design effective prompts, ground answers with retrieval, orches- trate multi-step workflows, and monitor how the system behaves once it’s deployed. My own journey into this field followed a similar path to many developers. I began by experimenting in Jupyter Notebooks, exploring APIs, and learning where the mod- els succeeded and where they struggled. Those early explorations gradually evolved into small, agent-based side projects at work, built for a handful of early adopters seek- ing productivity gains. As the technology advanced—through improvements in OpenAI’s APIs, the growing capabilities of LangChain, the orchestration power of LangGraph, and emerging techniques such as advanced Retrieval-Augmented Gener- ation (RAG) and ReAct—those prototypes became more sophisticated. At the same time, I began writing this book. More than once, a chapter I had just completed feltxvi outdated only a few weeks later. It was both exhilarating and, at times, exhausting. Licensed to THIAGO BANDEIRA <thiago@lar.ifce.edu.br>

PREFACE xvii That experience strongly influenced the way this book is written. Rather than chase every new parameter or short-lived “best practice,” the focus here is on the con- cepts, architectures, and design patterns that have proven stable and that underpin reliable LLM applications. We’ll build concrete systems—engines, chatbots, and agents—but the aim is to provide reusable foundations: how to structure retrieval, design prompts, compose chains, evaluate behavior, and orchestrate multi-step work- flows with clarity and confidence. Frameworks play a crucial role in this process. LangChain standardizes the essen- tial components—loaders, splitters, embeddings, retrievers, vector stores, and prompts—so that you don’t need to reinvent the plumbing for every project. Lang- Graph extends this by structuring workflows as graphs and coordinating agent loops, while LangSmith adds visibility for debugging and evaluation. Together, they enable developers to focus on the application itself rather than the underlying infrastructure. Why this book, and why now? The foundations are finally stable enough to teach effectively, and the need is greater than ever. Development teams want practical guid- ance that bridges the gap between ideas and implementation without tying them to a single vendor or a fleeting technique. If this book succeeds, you’ll come away with a clear mental model for building LLM-powered systems, a set of proven patterns you can depend on, and the confidence to keep building—even as the landscape contin- ues to evolve beneath our feet. Licensed to THIAGO BANDEIRA <thiago@lar.ifce.edu.br>

acknowledgments Writing a book like this is never a solo effort, and I feel profoundly grateful to have had so many talented and generous people alongside me on this journey. First, I would like to thank Juan-Mauro Bozzano, Romulus Corneanu, Carolyn Kao, and Ivan Lin for their thoughtful feedback at the very beginning, when the first chapters were still little more than outlines. Their comments and conversations helped shape the direction of the book in its early stages. I’m also indebted to David Bujan and Diego López-de-Ipiña from Deusto University, who encouraged me to bring greater preci- sion and rigor to the work—especially in those formative drafts when many of the ideas were still taking shape. A special word of thanks goes to Michael Stephens at Manning, who believed in this project from the very first conversation. At the time, LLMs were only beginning to capture the world’s attention, and the road ahead was far from clear. His support, together with timely guidance and encouragement to adapt as the technology evolved, helped keep the book relevant and focused. I’m particularly grateful to my editor, Dustin Archibald, whose steady guidance and thoughtful advice carried me through every stage of the writing process, and to Andrew Freed, whose sharp technical insights made the explanations clearer and more accurate. Many thanks also to technical editor Antowan Batts. Antowan is a financial systems analyst and applied economist with expertise in finance, data analy- sis, and global supply chains. He teaches economics and focuses on connecting analyt- ical insights to real-world business challenges. To the reviewers—Abdullah Al Imran, Abhishek Guha, Andres Mariscal, Antowan Malik Batts, Artur Guja, Aryan Jadon, Balaji Venkateswaran, Borko Djurkovic, Byronxviii Galbraith, Carnell Greenfield, David Caswell, David Jacobs, Derek Morgan, Felipe P. Licensed to THIAGO BANDEIRA <thiago@lar.ifce.edu.br>