AI¶

RAG vs Fine-Tuning vs Training from Scratch: The Real Costs

Introduction

A question comes up frequently when I work with companies on their AI projects: "Should we train our own model?". Or the slightly more advanced variant: "We want to fine-tune a model on our data".

Every time, I need to take a moment to explain what that actually means in practice. Because between training a model from scratch, fine-tuning it on your own data, or simply giving it context with a RAG, there is a world of difference. In cost, time, complexity, and above all in outcome.

In this article, I will try to lay things out simply. What is an AI model, how do you train it, what does it cost, when is it worth it, and most importantly why in 95% of cases you probably do not need to do either.

How to Optimize RAG: 8 Techniques with Measured Gains

You're probably optimizing in the wrong place

When a RAG isn't working well, here's what 90% of teams do: they change the prompt.

They rephrase the instructions, try different models, adjust the temperature. And sometimes it helps a little. But most of the time, that's not where the problem is.

Jason Liu, one of the most followed RAG experts, has a framing I find spot-on: "Before touching anything, reach 97% recall in retrieval."

97% recall means that in 97 out of 100 cases, the chunk containing the right answer is among the results you pass to the LLM. If you're not there, the best prompt in the world won't change a thing. The LLM cannot invent information that isn't in its context.

The real RAG optimization order is: measure first, then retrieval, then generation. Not the other way around. If you're not yet familiar with the basics of how RAG works, start there before optimizing any component.

Optimal RAG Chunking : 8 Strategies & Real Benchmarks

The chunking you're probably using is the worst one tested

Let me start with a result that surprised me when I first saw it.

Chroma Research published a benchmark comparing all common chunking strategies. They tested the default OpenAI Assistants parameters: 800 tokens, 400 tokens of overlap. Their verdict is unambiguous — it's the configuration with the lowest precision across all tests. 1.4% precision. Their exact comment: "particularly poor recall-efficiency tradeoffs".

These are the parameters tens of thousands of projects are using right now, often because it's what the LangChain or LlamaIndex quick start suggests.

Meanwhile, configurations 4x simpler (200 tokens, zero overlap) perform 3.7x better on precision.

Chunking is the decision most teams spend the least time on. And yet it's probably the one with the highest impact on your RAG quality.

Hybrid RAG : BM25 + Vector Search With +10% Recall

Your vector RAG is missing questions you don't even know about

It's a comment I hear often on RAG projects: "It works well in general, but sometimes it finds nothing on questions that seem straightforward."

Concrete example: "What is the ISO-27001 procedure for remote access?" → 0 relevant results.

Vector search encodes meaning. But when a query contains an exact identifier — a standard name, a product code, a domain acronym — semantic encoding fails completely.

This is what's called vocabulary mismatch. And it's the problem hybrid search solves.

Agentic RAG vs Classic RAG : Do You Really Need It?

Your RAG isn't enough anymore. Really?

It's everywhere. Agentic RAG. The future of RAG. The next evolution.

And as usual with AI trends, you get the feeling that if you haven't switched to Agentic RAG yet, you're already behind. That your "classic" RAG is outdated. That you need to rebuild everything from scratch.

Here's what I actually think: it's not that simple, and most projects don't need Agentic RAG.

But — and there's always a but — Agentic RAG solves real problems that classic RAG simply cannot. And if you run into those problems, you'll need it.

So in this article, I'll keep it straightforward: what Agentic RAG actually is, how it differs from classic RAG, and most importantly, how to decide whether you need it.

The 5 Most Common RAG Mistakes

Introduction

Since 2023, I've run about ten RAG projects myself and led another ten with teams. Some went very well, others less so, but we always tried to learn and correct course along the way. Looking back, I keep seeing the same mistakes, whether from myself early on, from clients, or from peers. These aren't technical mistakes (I cover those in this article), they're mistakes of posture, approach, and method.

Everyone makes them at least once. The point here is to name them clearly so they don't get repeated.

RAG vs Long Context LLM: Is RAG Really Dead?

Introduction: RAG, a magic method?

Every time a new model launches with a larger context window, people announce that RAG is obsolete. Yet RAG was born out of a very concrete need: you cannot hand a 400 or 500-page document to an LLM and ask it questions on the spot.

In enterprise settings, you often have dozens (or even hundreds) of files. RAG offers a simple answer: build a document base out of small pieces (chunks), then dynamically supply the relevant chunks to the AI with each question.

4 Technical Causes of RAG Failure (and How to Fix Them)

Introduction

A "basic" RAG is quick to set up, but it often plateaus between 50 and 70% correct answers. In a business context, that's not good enough for reliable use.

If you're looking for an error analysis method to prioritize improvement actions, the dedicated article is here:
My RAG isn't working: why error analysis changes everything

If you want to first understand why RAG remains useful despite large context windows, I have a dedicated article:
Is RAG really dead?

Here, we focus on the other question: why a RAG doesn't answer correctly, and how to improve it.

What Is an AI Agent? Clear Definition With Real Examples

The 2025 AI trend: AI agents

You may have thought: "Another new buzzword." Honestly, I get it.

A few months ago, everyone was talking about RAG — the AI that was supposedly going to revolutionize everything and replace entire workforces with knowledge bases. Now it's AI agents, presented as the inevitable next step.

In reality, this is yet another AI technology, and there's a push to convince you that you absolutely need it. For the record, I get along perfectly well without an AI agent that makes my coffee, cooks my meals, and tidies my apartment. But — and there's always a but — these AI agents genuinely solve real problems and address real business needs.

So, what exactly is an AI agent? What does agentic AI actually mean? To understand that, you first need to understand what ChatGPT is — and more importantly, what its limitations are. Because AI agents exist to address (or work around) the limitations of language models like ChatGPT, Gemini, and Claude.

What Is RAG? Definition, How It Works & Real Limits

Introduction to RAG (Retrieval-Augmented Generation)

Everyone has heard of RAG (Retrieval-Augmented Generation) at this point. But what is RAG exactly? Many people have already implemented it — sometimes with no-code tools or Python libraries like LangChain or LlamaIndex. It's straightforward to set up, but I also see a lot of people disappointed with the results. The thing is, you really need to understand what it's for and how it works before you can tell whether it's the right fit for your use case.

I hadn't originally planned to write another RAG explainer — there are already plenty of resources out there. But talking with people who want to use it in enterprise contexts, I keep noticing the same pattern: everyone rushes past the fundamentals. What does RAG actually do? How does it really work in practice?

So let me walk through the points I usually end up clarifying when someone asks me about it.