04 · Retrieval-Augmented Generation (RAG)

What you’ll learn

RAG = retrieve relevant documents, then let the model read them and answer. It’s the most widely used LLM pattern because it makes answers grounded (quoting real sources) and updatable (change the docs, not the model).

Run the demo and read the trace: you’ll see the retrieval step (with similarity scores) followed by the grounded generation step.

The pipeline

flowchart LR
  Q[Question] --> R[Retriever<br/>top-k by similarity]
  R --> C[Context chunks]
  C --> G[Generator<br/>answer from context]
  G --> A[Grounded answer]

1. Retrieve the top-k chunks for the question (Project 02’s machinery).
2. Augment the prompt by inserting those chunks as context.
3. Generate an answer grounded in that context.

import { ragPipeline, loadCorpus } from '@lib/js';

const corpus = await loadCorpus('knowledge-base', import.meta.env.BASE_URL);
const { answer, trace } = ragPipeline('What is retrieval augmented generation?', corpus);

// `trace` records each step: retrieve → answer, so you can inspect the mechanics.

from _shared.data import load_corpus
from rag import rag_pipeline

corpus = load_corpus("knowledge-base")
result = rag_pipeline("What is retrieval augmented generation?", corpus)
print(result["answer"])
for step in result["trace"]:
    print(step["step"], step["actor"], step["detail"])

Design choices that matter

• Chunk size. Too big wastes the context window; too small splits ideas. (See the Chunking note.)
• k. How many chunks to retrieve — balance recall vs. noise.
• Grounding instruction. Tell the model to answer only from the context to cut hallucination.

🔍 Want to go deeper? This project is Simple RAG — the baseline. Explore 18 RAG techniques (semantic chunking, reranking, fusion, HyDE, adaptive and corrective RAG) with animated pipelines and a benchmark leaderboard to see which upgrades actually pay off.

RAG answers from a single retrieval. When questions are vague or multi-part, you want the model to retrieve, judge, and retry — that’s Agentic RAG, next.