Generative AI, grounding and RAG
Business leaders do not need to build models, but they must know what generative AI can and cannot do. Most exam scenarios turn on choosing the right type of AI and keeping answers trustworthy.
Generative AI versus predictive AI
Generative AI creates new content: text, images, code, summaries and translations. It is the technology behind chatbots and Copilot. Predictive AI (machine learning) finds patterns in historical data to forecast outcomes, such as demand or customer churn.
| Need | Right fit |
|---|---|
| Draft a product description, summarize a call, translate text | Generative AI |
| Talk to customers and ask about their needs | Generative AI chatbot |
| Predict customer retention or forecast demand | Azure Machine Learning (predictive) |
| Read text from images or scanned forms | OCR or Azure Document Intelligence |
| Understand pictures | Computer vision |
A multi-modal solution handles more than one type of input, for example text plus images or blueprints.
Why answers can be wrong
A large language model (LLM) predicts likely text. It can produce confident but false statements, called hallucinations. It also only knows what it was trained on, so it does not know your company’s current documents.
Grounding
Grounding means anchoring a response in specific data sources, such as your SharePoint sites, a knowledge base or a database. A grounded answer is more accurate, more relevant and easier to verify. Microsoft 365 Copilot is grounded in Microsoft Graph: the emails, files, meetings and chats that the user is allowed to see.
Retrieval Augmented Generation (RAG)
RAG is the common way to ground a model. The steps are:
- The user asks a question.
- A search service, for example Azure AI Search, finds the relevant passages.
- Those passages are sent to the model with the question.
- The model writes an answer based on them.
The big benefit is currency without retraining. When a document changes, the next answer reflects it. Fine-tuning, by contrast, adjusts the model itself, costs more and does not keep up with changing content.
Prompts
Prompt quality drives answer quality. A strong prompt includes the goal, context, the sources to use and the expected format. Examples help, but a single vague requirement or heavy jargon does not.
Exam traps
- Grounding is about data sources, not about explaining how the model works (that is transparency).
- RAG is not fine-tuning. If the question says “without retraining”, think RAG.
- Machine Learning predicts; generative AI creates. A forecast question is not a generative AI question.
- Azure AI Search finds and retrieves information; it does not generate answers by itself.
- Content filtering blocks harmful output; it does not make answers accurate.
Practice questions
Choose an answer, then review the explanation to check your understanding.
1. A retailer wants an AI assistant that answers customer questions using only its verified product catalog. Which concept describes this requirement?
Show answer
Answer: B.Grounding anchors responses in specific, trusted data sources, which makes answers more accurate and relevant.
2. An assistant must use the latest internal policy documents without retraining the model each time they change. What should you recommend?
Show answer
Answer: A.RAG retrieves current content at question time and passes it to the model, so no retraining is needed.
3. Which task is the clearest example of a generative AI benefit?
Show answer
Answer: B.Generative AI creates new content such as text, summaries and translations. Forecasting is predictive AI.
4. Model answers are inconsistent. Which change to the prompt is most likely to help?
Show answer
Answer: B.Good prompts state the goal, context, sources and expected format.