Learning/AI Agents — Complete Guide/Lesson 13

Chapter 5·Lesson 1 of 3·10 min

RAG and knowledge

Learn retrieval-augmented generation, indexing, retrieval, ranking, grounding, and freshness.

Concept diagram

flowchart LR
D[Documents] --> C[Chunks]
C --> E[Embeddings]
E --> V[(Index)]
Q[Query] --> QE[Query representation]
QE --> V
V --> F[Permission filter + rerank]
F --> X[Grounded context]
X --> M[Model]
M --> A[Answer + sources]

Lesson overview

RAG and knowledge

Retrieval-augmented generation (RAG) supplies relevant external information to a model at run time. It is useful for private, changing or domain-specific knowledge that should not be assumed to live in model weights.

Ingestion

Clean documents, split them into meaningful chunks, attach metadata and create searchable representations such as embeddings. Preserve source, tenant, version, permission and update metadata.

Chunking

Very small chunks lose context; very large chunks reduce retrieval precision and consume the model context budget. Chunk by semantic boundaries where possible and test retrieval quality rather than choosing a universal size.

Query path

Transform the user's question into a search request, retrieve candidates, apply authorization filters, optionally rerank them and assemble a compact evidence set. Permission filtering must happen before content reaches the model.

Grounding

Preserve source IDs. For important answers, require citations or evidence references. Evaluate retrieval quality separately from answer quality so you can identify whether a failure came from search or generation.

When not to use RAG

If information is small and already available in context, retrieval may add unnecessary complexity. If the task requires transactional truth, query the authoritative application service instead of a stale document index.

The goal is not “use a vector database.” The goal is to supply the right evidence at the right time.

Learning path

Theory → Example → Code → Practice → Quiz → Challenge → Completion

0/6 done

Step 1

Theory

RAG and knowledge

Ingestion

Clean documents, split them into meaningful chunks, attach metadata and create searchable representations such as embeddings. Preserve source, tenant, version, permission and update metadata.

Chunking

Query path

Grounding

When not to use RAG

The goal is not “use a vector database.” The goal is to supply the right evidence at the right time.

Step 2

Example

Example: SaaS documentation assistant

Index documentation with product, version, tenant and source metadata. A query first establishes the user's access, retrieves candidate chunks, filters unauthorized documents, reranks the candidates and sends only the best evidence to the model with source IDs.

Step 3

Code

Permission-aware retrieval

typescript

const candidates = await vector.search({ embedding, filter: { tenantId } });
const allowed = candidates.filter(c => policy.canReadDocument(user, c.documentId));
const context = allowed.slice(0, 6).map(c => ({ text: c.text, sourceId: c.documentId }));

Retrieval must not become an authorization bypass.

Step 4

Practice

Design a RAG pipeline for a SaaS documentation assistant. Decide document boundaries, chunking, metadata, retrieval count, reranking, permission filtering, citation format and stale-document handling.

Step 5

Quiz

1. What is RAG primarily for?

2. When should document permissions be applied?

Step 6

Challenge

Create 20 evaluation questions where each correct answer requires a specific source document. Measure retrieval recall separately from final answer correctness.

Complete every stage

Work through every step in order, then the lesson will be marked complete.

Previous lessonLong-running tasks Next lessonThe retrieval pipeline

Each chapter and subtopic has its own public URL under /ai-agent.