LEARN · AI CONCEPTS · PRACTICE
22 AI Concepts Practice Questions
Four levels, LLM basics to combining RAG and MCP in one real system. These are scenario and design questions, not code, work through the reasoning, there's no single-line answer to check against.
Level 1 · LLM Basics
Q1–6Tokens, context windows, temperature, and prompting.
Tokens, Not Words
Explain why a model's context limit is measured in tokens rather than words, with a concrete example of a word that splits into multiple tokens.
Context Window Overflow
A conversation has grown past the model's context window. Describe what actually happens to the oldest messages, and one real strategy to handle it.
Temperature for a Support Bot
Would you set a high or low temperature for a customer-support bot that must give factually consistent answers? Justify it.
Temperature for Brainstorming
Would you set a high or low temperature for a creative brainstorming assistant? Justify it, contrasting with the previous question.
Rewrite a Vague Prompt
Rewrite a vague one-line prompt into a well-structured prompt that includes at least one concrete example (few-shot).
Why Bigger Context Isn't Always Better
Explain a real downside of simply maximizing an LLM's context window instead of retrieving only the relevant information.
Level 2 · RAG
Q7–12Embeddings, vector databases, chunking, and retrieval quality.
When RAG, Not Fine-Tuning
Describe a real scenario where RAG is clearly the better choice over fine-tuning a model, and explain why.
Embedding Similarity, in Plain Terms
Explain what it means for two pieces of text to have "similar embeddings," without using any math notation.
Choosing a Vector Database
List two real factors you'd weigh when picking a vector database for a RAG system with 10 million documents.
Chunking Strategy for Long PDFs
Design a chunking strategy for a knowledge base of 500-page technical manuals, explaining your chosen chunk size and overlap.
Chunking Strategy for Chat Logs
Would the same chunking strategy work for a knowledge base of short customer chat transcripts? Explain what you'd change.
Retrieval Gone Wrong
A RAG system keeps retrieving irrelevant chunks for a specific query. List two real causes, and how you'd diagnose which one it is.
Level 3 · MCP
Q13–18Clients, servers, tools, resources, and real workflows.
MCP Client vs Server, in Your Own Words
Explain the actual difference between an MCP client and an MCP server, using a concrete example of each.
Tool or Resource?
An MCP server exposes "get the current weather for a city." Is that a tool or a resource? Explain your reasoning.
Tool or Resource? (Part 2)
An MCP server exposes "the contents of a specific configuration file." Is that a tool or a resource? Explain your reasoning, contrasting with the previous question.
Designing a Simple MCP Workflow
Sketch the sequence of steps for an MCP client that needs to look up a customer's order status through an MCP server.
When an MCP Tool Call Fails
Describe what should happen when an MCP tool call returns an error, from both the server's side and the client's side.
One Server, Many Clients
Explain why the same MCP server can be reused by multiple different AI assistants without being rewritten for each one.
Level 4 · Putting It Together
Q19–22Deciding between RAG and MCP, and combining both in one real system.
RAG vs MCP, Which One?
You're building an assistant that needs to both answer questions from a document library and send real emails on request. Which parts need RAG, which need MCP, and why?
Combining RAG and MCP
Describe a single real application that would genuinely need both RAG and MCP working together.
The Real Difference
In one or two sentences, state the single clearest distinction between what RAG solves and what MCP solves.
Diagnosing the Wrong Tool Choice
An assistant was built using RAG to let a model "update a customer's shipping address," and it isn't working reliably. What's architecturally wrong here, and what should have been used instead?
- Write out the real reasoning, not just a one-word answer, these questions are testing judgment, not recall.
- Where two questions contrast (like the two temperature questions, or the two tool-or-resource questions), actually write both answers side by side, the contrast is the point.
- If you have access to an LLM, use it to check your reasoning against, but write your own answer first.