Augmentation is everything that happens between retrieval and generation: chunk ordering, grounding directives, and citation tagging. Reranking and query embedding are retrieval; fine-tuning changes the model.
Think of RAG as cooking dinner from a recipe box. Retrieval is fetching ingredients from the pantry: you embed the query and search the vector store. Generation is the chef cooking. Augmentation is everything in between: how you arrange the ingredients on the counter before the chef starts. You decide which order to lay them out so the most important one is not buried in the middle. You leave a sticky note saying 'only use what's on this counter, and label which item each dish came from.' That sticky note is your system prompt and your citation tags. Reranking is still pantry work: picking the best ingredients before they hit the counter. Fine-tuning is sending the chef to cooking school. None of those are arranging the counter, so none of them are augmentation.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Production RAG is usually drawn as three stages (retrieve, augment, generate), but in interviews the stage boundaries are where candidates slip. People can recite the pipeline yet cannot say cleanly which decisions belong to which stage, and that is exactly what a multi-select like this probes. The word 'augmentation' is the A in RAG, yet it is the stage most engineers describe vaguely, because the famous diagrams spend their detail on retrieval (embeddings, indexes, rerankers) and on the model, leaving the middle as a single arrow.
This question is a classification test. Six operations are listed; three are augmentation and three are not. The skill being measured is not whether you know what reranking or fine-tuning do in isolation, but whether you can place each operation on the correct stage boundary. The discriminating rule is simple to state and surprisingly easy to misapply under pressure: augmentation is the work of constructing the prompt from a finalized set of chunks. Anything that selects which chunks survive, or that changes the model itself, sits in a different stage.
We will work the rule first, then run all six options through it, then close on why interviewers reach for exactly this framing. By the end you should be able to defend each classification in one sentence apiece.
What augmentation actually owns
Augmentation is the middle stage of RAG, and its job is narrow but high leverage: take the chunks retrieval handed over and turn them into the exact prompt the model will read at inference. Nothing it does selects new chunks and nothing it does touches the model's weights. It only assembles text.
Three decisions live here. First, chunk ordering: where each surviving chunk lands in the context window. This is not cosmetic. Long-context models exhibit a positional bias where information in the middle of the context is recalled less reliably than information at the start or end. A serious augmentation step places the highest scoring chunk at an edge to fight that bias, and may interleave a short context-map header that tells the model what each numbered block is about.
Second, the system-prompt directives. These instruct the model to answer only from the supplied context, to cite the chunks it used, and to refuse or hedge when the context is insufficient. This is the single highest leverage knob for faithfulness in the entire pipeline. A two-line change to the grounding instruction (say, adding 'if the context does not contain the answer, reply that you do not know') often moves a faithfulness metric more than a week of index tuning, because it directly governs whether the model invents claims when retrieval comes up short.
Third, citation tagging: numbering chunks as [1], [2], and so on. That metadata is what lets the model attribute each claim to a source, and it is what a downstream eval harness uses to verify grounding. Without tags the output is a paragraph of text with no traceable provenance; with them, every sentence can be checked against the specific chunk it claims to rest on. That is why citation tagging is properly infrastructure, not decoration.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
# Stage boundaries in a RAG pipeline
# --- RETRIEVAL ---
qvec = embed(query) # query embedding (NOT augmentation)
cands = vector_store.search(qvec, k=20)
cands = reranker.rerank(query, cands)[:5] # reranking (NOT augmentation)
# --- AUGMENTATION ---
cands = order_for_attention(cands) # chunk ordering: best chunk at the edge
ctx = "\n".join(f"[{i+1}] {c.text}" for i, c in enumerate(cands)) # citation tagging
prompt = (
"Answer ONLY from the context. Cite chunks as [n]. "
"If context is insufficient, say you do not know.\n\n" # grounding directives
f"Context:\n{ctx}\n\nQuestion: {query}"
)
# --- GENERATION ---
answer = llm.generate(prompt)| Decision | Stage | Why |
|---|---|---|
| Chunk ordering in the prompt | Augmentation | Shapes the text the model reads; counters lost-in-the-middle |
| Grounding / refusal directives | Augmentation | System prompt governs how context is used |
| Citation tagging [1], [2] | Augmentation | Metadata in the prompt enabling attribution + eval |
| Cross-encoder reranking | Retrieval | Reorders candidates before the prompt is built |
| Query embedding | Retrieval | First step; triggers the vector lookup |
| Fine-tuning on grounded examples | Model training | Changes weights offline; orthogonal to the prompt |
Real products, models, and research that use this idea.
- Perplexity numbers its retrieved sources as inline citations and instructs the model to attribute each claim: citation tagging and grounding directives in action.
- Anthropic's Claude with the citations API tags document chunks so responses link claims back to source spans, an augmentation layer feature.
What an interviewer would ask next. Try answering before peeking at the approach.
QWhy does chunk ordering in the prompt change answer quality even when the same chunks are present?
Explain the lost-in-the-middle effect: long-context attention underweights the middle, so edge placement of the top chunk raises recall of that fact; cite ordering strategies like best at end.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Lumping reranking and query embedding into augmentation. Both are retrieval-stage work that runs before the prompt is assembled, not decisions about how context is presented to the model.
60 second bullets to scan on the way to the call.
The three RAG stages and what each one owns
Why chunk ordering is an augmentation decision, not cosmetic
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.