Level 4 — Expert

Imagine asking an AI about your company — and it makes up the answer because it doesn't know your internal documents. That's exactly the problem RAG (Retrieval Augmented Generation) solves.

The principle is elegant: before the AI answers, it first searches yourdata — documents, FAQs, product descriptions, internal wikis — and uses the relevant passages as context for its response.

In practice, RAG works in three steps: your documents are split into small chunks and stored as mathematical vectors (called embeddings). When a question comes in, it's also converted to a vector, and the most similar document chunks are found. These chunks are then sent to the AI along with the question as context.

The key to a good RAG system is the knowledge base. This is where you store everything the AI should know: company documents, product catalogues, FAQs, manuals, meeting notes, or customer feedback.

For this you need a vector database — a specialised database that doesn't search for exact words but for meaning. The concept behind it is called embeddings: every text is converted into a mathematical vector (a list of numbers) that represents its meaning. Similar texts have similar vectors.

The typical workflow: upload documents, split them into chunks, generate embeddings (e.g. with OpenAI's embedding models or Voyage AI), and store them in the vector database. When a query comes in, the most relevant chunks are retrieved and passed to the AI as context.

MCP(Model Context Protocol) is an open standard developed by Anthropic that defines how AI models communicate with external tools and data sources. Think of MCP as “USB for AI”: a universal connector that lets AI access any tool.

Before MCP, every tool had to build its own integration. With MCP, there's a standard — implement it once, and it works with all compatible AI systems. Since 2025, MCP has established itself as the defining standard, and an increasing number of tools and platforms support it.

What makes it special: the AI decides which tool to use and when. You say “Create a report on our latest sales figures,” and the AI knows it needs to first query the database, then analyse the numbers, and finally create a formatted report.

An MCP server is a small programme that sits between the AI and an external service. It provides “tools” that the AI can use. Setting one up is easier than you might think:

MCP is most commonly used with Claude (the desktop app or Claude Code). In the configuration file, you simply list your MCP servers, and Claude can automatically use their tools. The configuration is a simple JSON file.

You know the basics of prompting from the previous levels. Now it's about techniques that deliver professional results — reproducible and consistent.

Tip: combine these techniques. A system prompt with role definition plus few-shot examples plus structured output yields consistent, high-quality results — every time.

When to use which approach? This is a crucial question, and the answer depends on your specific use case. Here's an overview:

Rules of thumb:Always start with prompting. If the context window isn't enough or you have large amounts of documents: RAG. Fine-tuning only when you need a very specific style or domain language and the other approaches fall short.

When you build your own AI systems, you carry responsibility for the data you process. Especially in Germany and the EU, strict rules apply — and that's a good thing.

Important: never send sensitive data (passwords, credit card numbers, health data) to AI APIs without first checking whether the provider uses that data for training. Most providers don't when using the API — but verify it.