Chapter 12 — RAG (Retrieval-Augmented Generation)

TL;DR: RAG combines retrieval (embeddings + search) with generation (LLM). The agent retrieves relevant documents before generating the answer, reducing hallucinations and keeping responses grounded in facts.

The RAG pipeline

mermaid Copiar Tela cheia

graph TD
    Q[User question] --> E[Embed question]
    E --> S[Retrieve relevant chunks]
    S --> C[Build context]
    C --> P[Prompt with context + question]
    P --> LLM[LLM generates answer]
    LLM --> R[Fact-grounded response]

Código do diagrama

graph TD
    Q[User question] --> E[Embed question]
    E --> S[Retrieve relevant chunks]
    S --> C[Build context]
    C --> P[Prompt with context + question]
    P --> LLM[LLM generates answer]
    LLM --> R[Fact-grounded response]

RAG quality

Three metrics matter:

1. Recall — was the right chunk in the top-k retrieved?
2. Precision — how many retrieved chunks were actually relevant?
3. Latency — how long to retrieve, build context, and generate?

async function rag(query: string) {
  const chunks = await semanticSearch(query, k: 5);
  const context = chunks.map(c => c.content).join("\n---\n");

  return await llm.generate({
    prompt: `Based on this context:\n${context}\n\nAnswer: ${query}`
  });
}