2.3. Applying Generative AI Concepts

💡 First Principle: A foundation model starts with broad general knowledge and no knowledge of your business, your current facts, or your standards. Every technique in this section adds context or shapes behavior: prompts tell the model what you want right now, the context window limits how much it can consider at once, retrieval supplies current company facts, and fine-tuning changes how it behaves by default. Choose the cheapest technique that closes the gap between what the model knows and what the task needs.

Why care? Because most disappointing generative AI pilots fail for reasons this section explains: vague prompts, overstuffed context, a model asked about information it was never given, or an expensive fine-tuning project where retrieval would have worked. These are also cost decisions: every token costs money, and adaptation techniques range from free (better prompts) to substantial (fine-tuning and custom training).

The mental model is briefing a brilliant contractor who has just arrived. Prompting is the briefing: role, task, constraints, examples. The context window is the size of the desk: only so many documents fit at once. RAG is handing them the right files from your archive for each task. Fine-tuning is putting them through your company's training program so your style and procedures become second nature.

⚠️ Common Misconception: "Fine-tuning is how you give a model your current documents," "a bigger context window means you can include everything without cost," and "prompt wording barely matters." In reality, retrieval supplies current facts, context has cost and attention limits, and prompting is the first and cheapest lever.