RAG Fundamentals

The flow

1. 1.Ingest & chunk
2. 2.Embed & index
3. 3.Authorized retrieval
4. 4.Grounded answer
5. 5.Evaluate

Two pipelines, one evidence chain

Ingestion parses documents, chunks text and stores vectors with source/version metadata. At query time, authorize access, retrieve passages and ask the model to answer using the available evidence. RAG does not retrain model weights.

Keep document updates and deletions reflected in the index. A successful retrieval from a stale policy can still produce an incorrect business answer.

Abstention is useful behavior

If there is no relevant authorized evidence, return a known unsupported state instead of fabricating a source. Similarity thresholds are dataset-dependent; the fixture uses a threshold only to demonstrate the branch.

A citation ID is not enough: verify that the cited passage supports the claim. Retrieved instructions remain untrusted data. The application must enforce access controls and safe tool execution.

Inspect the right failure layer

Measure whether the right source appears in top-k, then whether the answer faithfully uses it. Poor retrieval requires chunking, indexing or query work; a larger generation model may not fix it. Compare a simple baseline before adopting advanced RAG.

Try it

import assert from 'node:assert/strict';
const documents = [
  { id: 'policy-A', tenant: 'A', text: 'Refunds are available within 14 days.', score: 0.9 },
  { id: 'policy-B', tenant: 'B', text: 'Confidential enterprise policy.', score: 1 },
];
function retrieve(tenant, minimumScore) {
  return documents.filter(d => d.tenant === tenant && d.score >= minimumScore);
}
function composeContext(tenant, minimumScore) {
  const sources = retrieve(tenant, minimumScore);
  return sources.length
    ? { status: 'READY', context: sources.map(s => `[${s.id}] ${s.text}`), citations: sources.map(s => s.id) }
    : { status: 'ABSTAIN', context: [], citations: [] };
}
assert.deepEqual(composeContext('A', 0.8).citations, ['policy-A']);
assert.equal(composeContext('A', 0.95).status, 'ABSTAIN');
assert.equal(composeContext('unknown', 0).status, 'ABSTAIN');
console.log(composeContext('A', 0.8));
// Fixed retrieval scores and context only: this example does not call an LLM.

Run with Node.js 22: node examples/generative-ai/rag-fundamentals.mjs. This is a deterministic teaching example, not a live model call.

Interview questions

Official references

• Spring AI — RAG

Continue learning

• LLM Evaluation →
• Tool Calling →