How does citation linkage between retrieved chunks and the model output rely on context-engineering choices?

Attempt it

Each retrieved chunk gets a stable identifier (chunk_id or short tag like [S3]) attached as a header inside the context block. The system instruction tells the model to cite by that identifier when it uses a chunk. The identifier survives reranking, trimming, and reordering because it is bound to the chunk text, not to its position. At read time the renderer maps the cited identifier back to the original source row. The fragile alternative is asking the model to invent citations from the chunk text, it confabulates URLs and titles. The robust pattern is explicit, position-independent identifiers placed inline with the chunks.

Your answer

Quick reference

TL;DR

Attach a stable id (chunk_id or short tag like [S3]) inline with each chunk, tell the model to cite by that id, and map ids back to sources at render time, so citations survive reranking and trimming.

Memory aid

Sign in for the mnemonic that makes this stick. See Pro pricing.

Easy to grasp

Imagine you give a kid a stack of recipe cards and ask which card they used to make dinner. If every card has a sticker on the corner that says S1, S2, S3, the kid can just say 'S2' and you know exactly which recipe. If the cards have no stickers, the kid has to describe the recipe back to you, and they will probably get the title slightly wrong. RAG citations work the same way. Give every chunk a sticker before you hand them to the model, tell the model to cite the sticker, and you always know which chunk the answer came from. No stickers and the model has to guess at titles and URLs, and it guesses badly.

Key concepts

Context engineering Retrieval-augmented generation

Quick scan · 2–4 min

Everything important, quickly.

5 minutes: stable-id assembly pattern, why position-based numbering breaks, render-time mapping, name structured-citation APIs (Anthropic, OpenAI Responses, Cohere), and the hallucinated-tag validation step.

Code example

python

def assemble_context(chunks: list[Chunk]) -> str:
    blocks = []
    for i, c in enumerate(chunks, start=1):
        tag = f"S{i}"
        header = f"[{tag}] Source: {c.source_path} | section: {c.section}"
        blocks.append(f"{header}\n{c.text}")
    return "\n\n".join(blocks)

SYSTEM = (
    "When you use information from a labelled source block, cite it inline "
    "using its tag in square brackets (e.g. [S1]). Cite every claim that "
    "depends on a source. Do not invent tags."
)

# Render-time validation
def validate_citations(answer: str, valid_tags: set[str]) -> list[str]:
    import re
    cited = set(re.findall(r"\[(S\d+)\]", answer))
    return list(cited - valid_tags)  # hallucinated tags

Real-world examples

Real products, models, and research that use this idea.

Anthropic Claude Opus 4.7 ships a structured citation API: chunks are passed as `documents` and the model emits citation objects with stable doc ids.
OpenAI's 2026 Responses API exposes a document id based citation field for GPT-5, replacing earlier ad-hoc citation prompting.
Cohere Command R+ is trained on a `documents` parameter so citation ids round-trip cleanly through the chat completion API.
Perplexity Pro renders inline [1], [2] tags in the answer that resolve to the underlying source rows pulled at retrieval time.
LlamaIndex's CitationQueryEngine assembles per-chunk id headers and validates emitted citations against the retrieved set.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QHow do you handle citation linkage when chunks are summarized before injection?

Sign in or go Pro to see the approach for this follow-up.

QWhat changes when the model is asked to synthesize a fact across two chunks?

Sign in or go Pro to see the approach for this follow-up.

QHow do you catch a model that emits a tag not in the assembled context?

Sign in or go Pro to see the approach for this follow-up.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

Sign in for 5 unlocks/day or go Pro for unlimited revision sheets. See Pro pricing.

References

Primary sources. Skim if you want the original framing.