What is RAG and why use it instead of fine-tuning for knowledge?

// RAG vs "knowledge baked into weights", with deterministic stubs (no model is called).
// The retriever is a tiny keyword index; the "generator" answers only from retrieved text.
const STOP = new Set("a an the is are do does i my of to in on for what how many can you we our with be it".split(" "));
const stem = (w) => (w.length > 4 && w.endsWith("ed") ? w.slice(0, -2) : w.length > 3 && w.endsWith("s") ? w.slice(0, -1) : w);
const words = (s) => s.toLowerCase().replace(/[^a-z0-9 ]/g, " ").split(/\s+/).filter((w) => w && !STOP.has(w)).map(stem);

const docs = new Map(); // id -> { tenant, text }  (stands in for a vector store)
const upsert = (id, tenant, text) => docs.set(id, { tenant, text });

function retrieve(question, tenant, k = 2) {
  const q = new Set(words(question));
  return [...docs.entries()]
    .filter(([, d]) => d.tenant === tenant) // scope: a tenant only sees its own docs
    .map(([id, d]) => ({ id, text: d.text, score: words(d.text).filter((w) => q.has(w)).length }))
    .filter((h) => h.score > 0)
    .sort((a, b) => b.score - a.score || (a.id < b.id ? -1 : 1))
    .slice(0, k);
}

function buildPrompt(question, hits) {
  const ctx = hits.map((h) => `[${h.id}] ${h.text}`).join("\n");
  return `Answer only from the sources. Cite the source id. If the sources do not answer, say you do not know.\nSources:\n${ctx}\nQuestion: ${question}`;
}

// Stub generator: copies the best-matching source sentence and cites it; refuses when nothing matches well.
function generate(question, hits) {
  const best = hits[0];
  if (!best || best.score < 2) return "I do not know based on the provided documents.";
  return `${best.text} [${best.id}]`;
}

// Stub "fine-tuned" model: facts memorised at training time, no sources, never says "I do not know".
const trainedSnapshot = {
  "return": "You have 30 days to return an item.",
  "ship": "We ship to every country within 5 business days.",
};
function fineTuned(question) {
  const ws = words(question).join(" ");
  if (ws.includes("return")) return trainedSnapshot.return;
  return trainedSnapshot.ship; // closest memory, stated confidently
}

const ask = (label, question, tenant = "acme") => {
  const hits = retrieve(question, tenant);
  console.log(`${label}`);
  console.log(`  retrieved : ${hits.map((h) => `${h.id}(${h.score})`).join(", ") || "(none)"}`);
  console.log(`  RAG       : ${generate(question, hits)}`);
  console.log(`  fine-tuned: ${fineTuned(question)}`);
};

upsert("returns-v1", "acme", "Items can be returned within 30 days of delivery for a full refund.");
upsert("shipping", "acme", "We ship to the UK and the EU; delivery takes 3 to 5 business days.");
upsert("warranty", "acme", "Electronics carry a 2 year warranty covering manufacturing defects.");
upsert("returns-b", "globex", "Globex accepts returns within 60 days of delivery.");

console.log("== day 1: both systems know the return window ==");
ask("Q: How many days do I have to return an item?", "How many days do I have to return an item?");

console.log("\n== day 2: policy changes to 14 days; RAG re-indexes one document ==");
docs.delete("returns-v1");
upsert("returns-v2", "acme", "Items can be returned within 14 days of delivery for a full refund.");
ask("Q: How many days do I have to return an item?", "How many days do I have to return an item?");

console.log("\n== a question the documents do not answer ==");
ask("Q: Do you ship to Mars?", "Do you ship to Mars?");

console.log("\n== same question, another tenant (metadata filter) ==");
const hb = retrieve("How many days do I have to return an item?", "globex");
console.log(`  globex RAG: ${generate("x", hb)}`);

console.log("\n== the prompt the generator actually receives ==");
const q = "How many days do I have to return an item?";
console.log(buildPrompt(q, retrieve(q, "acme")));

// Cost of a knowledge update, illustrative: re-embed 1 doc vs a fine-tuning run over the whole corpus.
const corpusTokens = 2_000_000, docTokens = 400, epochs = 3;
console.log("\n== tokens processed to apply one policy change (illustrative) ==");
console.log(`  RAG re-embed one doc : ${docTokens.toLocaleString("en-US")} tokens`);
console.log(`  fine-tune, ${epochs} epochs : ${(corpusTokens * epochs).toLocaleString("en-US")} tokens (${(corpusTokens * epochs / docTokens).toLocaleString("en-US")}x)`);