Why should you care?

Company policies and product docs change weekly; retrieval keeps answers fresh without fine-tuning.

See it live — copy this example

Use .NET 8+. Run Semantic Kernel samples in a console or Web API. Keep API keys in user secrets or environment variables — never in source.

async Task<string> AnswerWithRagAsync(string question, IRetriever retriever, IChatClient chat)
{
    var chunks = await retriever.SearchAsync(question, topK: 4);
    var context = string.Join("\n---\n", chunks);
    var prompt = $"Use ONLY this context. If missing, say you do not know.\n{context}\nQ: {question}";
    return await chat.CompleteAsync(prompt);
}

What happened?

• Retriever returns text chunks.
• The prompt forbids outside knowledge.
• The chat client (SK or Extensions.AI) completes the answer.

Practice next

1. build a fake retriever with 2 hard-coded paragraphs.
2. Ask a question covered by the text.
3. Ask a question not covered and expect “do not know”.
4. Try topK=2 vs 6.
5. Add a tool wrapper SearchDocs(query).

Remember

Retrieve then generate. topK controls context size. Cite or log chunk ids.

Policy agent

HR asks leave policy questions.

Outcome: Answers grounded in handbook chunks.