Decoding ML is an educational newsletter that provides content on designing, coding, and deploying production-grade AI systems with software engineering and MLOps best practices to help you ship AI applications.

Our motto is to learn production AI by doing!

Thus, in addition to the newsletter, we offer five free courses on building end-to-end AI applications. If you thrive on hands-on experiences and building projects, these courses are for you.

One example is the Second Brain AI Assistant open-source course, which comprises six modules that explore advanced techniques, including agentic RAG, fine-tuning LLMs, and LLMOps.

To create your production-ready AI assistant, you’ll connect all the dots by building modular pipelines for data, features, training, inference, and observability.

Welcome to this week's AI roundup! We’re seeing OpenAI's ambitious expansion plans, Apple's daring exploration of a potential search engine shift, and Amazon's Vulcan robot bringing delicate handling to automation; the AI landscape is evolving rather rapidly. In this issue, we also have expert insights from Dr. Ben Auffarth on integrating RAG agents with LangGraph. Sold?

First up, our top stories of the week.

LLM Expert Insights,

Packt

Wysh Life Benefit allows any financial institution to offer free life insurance directly through their customers’ savings accounts. By embedding micro life insurance into deposit accounts, Life Benefit provides built-in financial protection that grows with account balances. It’s a simple, no-cost innovation that enhances loyalty, encourages deposits, and differentiates institutions in a competitive market. No paperwork. No medical exams. Just automatic coverage that provides peace of mind—without changing how customers bank.

Yes, absolutely! says Dr. Ben Auffarth, Chief Data Officer, Chelsea AI Ventures Ltd, and author of Generative AI with LangChain, published by Packt.

To understand why RAG is far from obsolete, let’s take a look at the patterns Ben highlights and and how LangGraph makes them more powerful.

LangGraph enables the creation of graph-based applications where runnables (i.e., composable units like chains, tools, or language model calls) act as nodes, and transitions between them serve as edges. It supports persistent state management, particularly useful for handling cyclical flows and maintaining context in multi-turn conversations for typical RAG systems. This persistent state allows the system to retain and evolve context over time. Thus, by facilitating decision-making based on intermediate results, LangGraph empowers RAG workflows to dynamically adjust their paths based on prior outcomes.

Ben identifies three advanced RAG patterns that take full advantage of this flexibility in his blog. Let’s look at them.

1. Conversational memory for RAG

One of the key challenges in

RAG is follow-up questions in multi-turn conversations,especially when users leave out critical context. LangGraph addresses this issue through stateful conversation management. In LangGraph, the conversation state (history of user and assistant messages) is maintained. This state becomes an input for nodes (runnables), enabling query rewriting, where the current user query can be augmented based on historical context. This allows for more targeted and context-aware retrieval, ensuring that the RAG system retrieves information relevant not just to the current question but to the entire conversation thread.

2. Hybrid retrieval with knowledge graphs

RAG systems need to capture information fromboth structured and unstructured sources to effectively augment model responses. RAG can perform vector searches to identify relevant documents, articles, etc. To capture facts and relationships between entities, however, a RAG system needs to work with structured knowledge bases like knowledge graphs.

Knowledge graphs are extremely useful as they represent entities as nodes and their relationships as edges, making it easier to capture and query complex

relationships. LangGraph enables hybrid retrieval by combining both vector searches and graph queries, leading to more semantically rich and factually grounded outputs.

3. Agentic RAGs

As AI agents become more capable, RAG systems must keep up—handling complex reasoning and dynamic decision-making, including interpreting queries, planning multi-step retrieval strategies and refining search queries iteratively. A popular approach that facilitates this dynamic retrieval strategy is a ReAct (Reasoning + Acting) loop. In ReAct, an agent interleaves reasoning steps (like language model-generated planning) with actions (e.g., querying a retriever, calling a tool, or accessing an API). This loop allows the system to decompose complex queries, determine what to retrieve and when, and refine or redirect the retrieval strategy based on intermediate observations. There’s much more to uncover about how LangChain and LangGraph can supercharge your RAG systems.

Grab a copy of Generative AI with Langchain, Second Edition written by Ben Auffarth and Leonid Kuligin.

Build production ready LLM applications and advanced agents using Python and LangGraph.

Join Packt’s Accelerated Agentic AI Bootcamp this June and learn to design, build, and deploy autonomous agents using LangChain, AutoGen, and CrewAI. Hands-on training, expert guidance, and a portfolio-worthy project—delivered live, fast, and with purpose.

🎓 As part of the exclusive Packt community, you get 50% off with code EXCLUSIVE50. Limited seats available.

📢 If your company is interested in reaching an audience of developers and, technical professionals, and decision makers, you may want toadvertise with us.

If you have any comments or feedback, just reply back to this email.

Thanks for reading and have a great day!

That’s a wrap for this week’s edition of AI_Distilled 🧠⚙️

We would love to know what you thought—your feedback helps us keep leveling up.

Thanks for reading,

The AI_Distilled Team
(Curated by humans. Powered by curiosity.)