Demystifying LLM-Based AI: A Practical Guide to Key Concepts

If you’ve ever used ChatGPT or another AI tool and wondered how it actually works, you’re not alone. At McKenna Consultants, we’re now incorporating AI into the bespoke software solutions we build for our clients. As part of that journey, we’ve found it helpful to explain the core concepts behind modern AI systems — particularly those based on Large Language Models (LLMs) — in simple terms.

This post takes you on a journey from the foundational ideas to more advanced terminology, demystifying the technology that powers tools like ChatGPT and the AI systems we’re building today.

What Is an LLM?

LLM stands for Large Language Model — a type of artificial intelligence trained to understand and generate human-like text. Think of it as an advanced autocomplete on steroids. LLMs are trained using enormous datasets of text (books, articles, websites, etc.), which they use to learn patterns in language, grammar, logic, and even style.

Popular LLMs include OpenAI’s GPT models, Anthropic’s Claude, Meta’s LLaMA, and Google’s Gemini.

Neural Nets: The Brains Behind the Model

The core of an LLM is a neural network, a kind of computer architecture inspired by the human brain. Neural networks consist of layers of artificial “neurons” that process input and learn from it.

In simple terms:

• You feed in a word (or part of a word).
• The neural network analyzes it and predicts what should come next.
• It does this using weights and heuristics learned during training — essentially, a giant math equation fine-tuned over time to get better at guessing what comes next.

This prediction-based mechanism is what allows an LLM to write poetry, explain a legal document, or help you debug code.

Context: The Conversation Window

LLMs operate based on context — the amount of text they can “see” at once when generating a response. Think of it like the AI’s short-term memory.

For example:

• GPT-3.5 can typically handle about 4,000 tokens (~3,000 words).
• GPT-4 Turbo can go up to 128,000 tokens (~96,000 words).

That context window determines how much information the model can consider when formulating a reply — which matters a lot in complex or multi-turn conversations.

Applications: Putting LLMs to Work

An LLM application is any tool that wraps an LLM in a user-friendly interface. ChatGPT, GitHub Copilot, AI-powered email assistants — all are examples of applications built on top of LLMs.

In our work at McKenna Consultants, we’re currently building an AI product assistant for an eCommerce website — a custom application that leverages LLM capabilities to improve the customer experience.

Agents: Giving AI a Sense of Autonomy

An AI agent is an LLM (or set of models) paired with memory, goals, and tools. Instead of just answering one question, agents can plan, reason, and take multiple steps toward a goal.

Imagine asking an agent to:

“Compare the prices of similar products across five competitor websites and give me a summary.”

An agent might:

1. Use web browsing tools.
2. Extract product information.
3. Organize the data.
4. Present the results.

It’s not just generating text anymore — it’s acting.

Tools: Extending the Model’s Reach

Most LLMs are limited to their training data and can’t fetch real-time info. But when paired with tools — like calculators, web browsers, databases, or APIs — they become far more powerful.

In our own implementations, we often give an AI system access to a company’s internal tools (like product databases or support documentation) to enhance its usefulness.

RAG: Retrieval-Augmented Generation

RAG is a technique that combines LLMs with external data sources. Instead of asking the LLM to “remember” everything, we store knowledge in a searchable database. When a user asks a question, the app:

1. Retrieves relevant info.
2. Feeds it into the model as context.
3. Generates an accurate, up-to-date response.

This is crucial for keeping AI helpful in domains with constantly changing information — like product specs or regulatory compliance.

Agent2Agent: Collaboration Between AIs

In more advanced systems, you might have multiple agents working together. For example, one agent might be responsible for research, while another writes the report. This approach is known as Agent2Agent communication.

This architecture is useful in workflows where different tasks require different kinds of expertise or models — and we’re just starting to explore what’s possible here.

Model Context Protocols: Organising the Conversation

As systems grow more complex, engineers need a way to structure how LLMs talk to users, tools, and each other. That’s where Model Context Protocols come in — they define how to format input and output, manage context, and coordinate between agents.

Think of it as a shared language that helps all the parts of an AI system stay on the same page.

Wrapping Up

Understanding the fundamentals of LLM-based AI helps make sense of the amazing things these systems can do — and their limitations. At McKenna Consultants, we’re blending this cutting-edge technology with our long-standing expertise in bespoke software development to deliver smarter, AI-enhanced solutions.

If your business is exploring how to integrate AI into your software products — whether that’s through chatbots, automation, or something more ambitious — we’d love to help.

Let’s build something intelligent, together.

To find our more about our AI development services, get in touch today.