RAG Fundamentals — Core

The retrieve-then-generate pipeline: loading, chunking, embedding, retrieval, reranking, grounded generation, evaluation, hybrid search, and RAG vs fine-tuning.

Medium50 questions

Sample questions

What problem does Retrieval-Augmented Generation primarily solve?

It permanently updates the model's weights with new facts at query time
It grounds answers in retrieved external documents, not parametric memory alone
It compresses the model so it runs on smaller hardware
It removes the need for a prompt by inferring intent from embeddings

Why

RAG retrieves relevant text at query time and conditions generation on it, grounding answers in an external, updatable knowledge source rather than only the model's frozen parameters. It doesn't change weights, shrink the model, or replace the prompt.

Retrieval-Augmented Generation (Lewis et al.)RAG for LLMs: A Survey

What is the correct high-level order of a basic RAG pipeline at query time?

Generate a draft, then retrieve documents to verify it
Fine-tune on the query, then generate from the updated weights
Rerank the whole corpus, then embed only the top document
Embed the query, retrieve relevant chunks, then generate grounded on them

Why

At query time you embed the user's query, retrieve the most relevant chunks from the index, and generate an answer conditioned on them. Generating first, fine-tuning per query, or reranking the entire corpus are not the standard flow.

RAG for LLMs: A Survey Retrieval-Augmented Generation (Lewis et al.)

Why are documents split into chunks before indexing in a RAG system?

To reduce the model's parameter count during retrieval
To retrieve focused, well-embedding passages that fit the context
To convert the documents into a single averaged embedding
To guarantee the whole document is always returned intact

Why

Chunking gives retrieval the right granularity: small enough to be a focused, well-embedded unit and to fit the context budget, but big enough to stay coherent. It doesn't affect parameters, isn't a single averaged vector, and deliberately avoids returning whole documents.

RAG for LLMs: A Survey Anthropic: Contextual Retrieval

What is the purpose of overlap between adjacent chunks?

To store each chunk twice for redundancy against data loss
To keep boundary-straddling context attached to its sentence
To reduce the total number of embeddings that must be stored
To force every chunk to have exactly the same token count

Why

A little overlap keeps a sentence or idea that spans a chunk boundary intact in at least one chunk, so retrieval doesn't miss context cut in half. It isn't for redundancy or uniform sizing, and it increases (not reduces) the number of chunks.

RAG for LLMs: A Survey Anthropic: Contextual Retrieval

In dense retrieval, how does the system decide which chunks are relevant to a query?

By comparing the query embedding to chunk embeddings via a similarity metric
By counting shared keywords between the query and each chunk
By asking the LLM to read every chunk in the corpus first
By selecting the most recently added chunks by timestamp

Why

Dense retrieval embeds the query and chunks into the same vector space and ranks by similarity (e.g. cosine/dot product), capturing meaning beyond exact words. Keyword counting is lexical (BM25), reading the whole corpus is infeasible, and recency isn't relevance.

Dense Passage Retrieval (DPR)Sentence-BERT

Free account

Take the full module

These are the first few of 50 questions. A free account opens the rest as a scored drill.

Every question in this module
Instant feedback and supporting reading
Your score and progress, saved

Free · your email is used for progress only.

About Nybble^™

The AI space moves fast.
Nybble^™ is how you keep up — and stay sharp.

Snack

What happened. In two minutes.

The AI news cycle moves at a pace no one can keep up with. Snack distills what launched, what shipped, and what matters — every day, without the filler.

Go to Snack

Stack

The concepts behind the headlines.

News tells you what. Stack tells you why and how. From RAG architectures to agentic evals, these are the ideas that will shape what you build next.

Go to Stack

Hack

Prove you actually get it.

Reading about LangChain is not the same as knowing it. Hack challenges you with production-grade questions, then shows you the references that make the answer stick.

Go to Hack