20. RAG Interview Questions and Answers
30 questions covering ingestion, chunking, embeddings, search, ranking, grounding, security, evaluation, and production.
Interview Q01: What is RAG?
Retrieves external evidence at query time and supplies it to a model; it is not a guarantee of truth.
Interview tip: define the concept, explain a trade-off, and name a production control or metric.
function validateCitation(id, allowedIds) {
if (!allowedIds.has(id)) throw new Error("Invalid citation");
return id;
}- Inspect the relevant pipeline stage instead of guessing.
- Preserve source provenance, versions, and access scope.
- Measure quality on representative queries.
What does RAG add?
Interview Q02: When should you avoid RAG?
Use direct database queries for exact transactional values and deterministic code for arithmetic.
Interview tip: define the concept, explain a trade-off, and name a production control or metric.
function validateCitation(id, allowedIds) {
if (!allowedIds.has(id)) throw new Error("Invalid citation");
return id;
}- Inspect the relevant pipeline stage instead of guessing.
- Preserve source provenance, versions, and access scope.
- Measure quality on representative queries.
Which is a direct database query?
Interview Q03: Name the main RAG stages.
Ingestion, parsing, chunking, indexing, retrieval, ranking, context assembly, generation, validation, and monitoring.
Interview tip: define the concept, explain a trade-off, and name a production control or metric.
function validateCitation(id, allowedIds) {
if (!allowedIds.has(id)) throw new Error("Invalid citation");
return id;
}- Inspect the relevant pipeline stage instead of guessing.
- Preserve source provenance, versions, and access scope.
- Measure quality on representative queries.
Which occurs before query-time retrieval?
Interview Q04: RAG vs fine-tuning?
RAG supplies context at inference; fine-tuning changes model behavior through training.
Interview tip: define the concept, explain a trade-off, and name a production control or metric.
function validateCitation(id, allowedIds) {
if (!allowedIds.has(id)) throw new Error("Invalid citation");
return id;
}- Inspect the relevant pipeline stage instead of guessing.
- Preserve source provenance, versions, and access scope.
- Measure quality on representative queries.
Which suits changing private docs?
Interview Q05: What is chunking?
Dividing documents into coherent retrievable units while preserving provenance.
Interview tip: define the concept, explain a trade-off, and name a production control or metric.
function validateCitation(id, allowedIds) {
if (!allowedIds.has(id)) throw new Error("Invalid citation");
return id;
}- Inspect the relevant pipeline stage instead of guessing.
- Preserve source provenance, versions, and access scope.
- Measure quality on representative queries.
What trade-off matters?
Interview Q06: Why overlap chunks?
Overlap preserves boundary context, but too much overlap creates duplicates and overhead.
Interview tip: define the concept, explain a trade-off, and name a production control or metric.
function validateCitation(id, allowedIds) {
if (!allowedIds.has(id)) throw new Error("Invalid citation");
return id;
}- Inspect the relevant pipeline stage instead of guessing.
- Preserve source provenance, versions, and access scope.
- Measure quality on representative queries.
Excessive overlap causes?
Interview Q07: What are embeddings?
Learned vector representations used for similarity search.
Interview tip: define the concept, explain a trade-off, and name a production control or metric.
function validateCitation(id, allowedIds) {
if (!allowedIds.has(id)) throw new Error("Invalid citation");
return id;
}- Inspect the relevant pipeline stage instead of guessing.
- Preserve source provenance, versions, and access scope.
- Measure quality on representative queries.
What do embeddings provide?
Interview Q08: Cosine vs dot product?
Cosine compares direction; raw dot product is magnitude-sensitive unless normalized.
Interview tip: define the concept, explain a trade-off, and name a production control or metric.
function validateCitation(id, allowedIds) {
if (!allowedIds.has(id)) throw new Error("Invalid citation");
return id;
}- Inspect the relevant pipeline stage instead of guessing.
- Preserve source provenance, versions, and access scope.
- Measure quality on representative queries.
Which compares direction?
Interview Q09: What is BM25?
A lexical ranking function using term frequency, inverse document frequency, and document length.
Interview tip: define the concept, explain a trade-off, and name a production control or metric.
function validateCitation(id, allowedIds) {
if (!allowedIds.has(id)) throw new Error("Invalid citation");
return id;
}- Inspect the relevant pipeline stage instead of guessing.
- Preserve source provenance, versions, and access scope.
- Measure quality on representative queries.
BM25 is strong for?
Interview Q10: What is hybrid search?
Combines lexical and vector retrieval to use exact and semantic matching.
Interview tip: define the concept, explain a trade-off, and name a production control or metric.
function validateCitation(id, allowedIds) {
if (!allowedIds.has(id)) throw new Error("Invalid citation");
return id;
}- Inspect the relevant pipeline stage instead of guessing.
- Preserve source provenance, versions, and access scope.
- Measure quality on representative queries.
Why combine them?