Skip to content

Module 6 min

The RAG Pipeline

Retrieve → rerank → synthesize → cite. Win at every stage or you never reach the answer.

  1. The RAG Pipeline
  2. 1Retrieval via embeddings
  3. 2Reranking signals
  4. 3The synthesis & citation step

Why it mattersKnowing the stages tells you exactly where a missing citation was lost.

01

Definition & Foundation

What it is, in plain words

RAG (retrieval-augmented generation) is the assembly line inside every major answer engine: retrieve → rerank → synthesize → cite. The engine fetches candidate passages that semantically match the query, re-scores that shortlist on authority, freshness, and answer-fit, hands the survivors to the model, and the model writes the answer while crediting a few sources.

The reason this diagram is worth memorizing: it turns "why aren't we cited?" from a mystery into a diagnosis. A missing citation was lost at a specific stage, and each stage fails differently: not retrieved means a content or access problem; retrieved but cut means an authority or freshness problem; surviving but unquoted means your phrasing wasn't liftable. Win at every stage or you never reach the answer; find the failing stage and you know exactly what to fix.

02

The Four Stages

What happens at each, and what failure looks like there

are you considered?

1 · Retrieve

Your pages are chopped into chunks, embedded as vectors, and matched to the query by meaning, not keywords. Failure here looks like total absence: you never appear for prompts you should own. Fix: retrievable, self-contained chunks (Retrieval & Embeddings) and crawler access (AI Crawlers).

do you survive?

2 · Rerank

The shortlist is re-scored on authority, freshness, and answer-fit; most retrieved passages are cut here. Failure looks like near-misses: you surface for related prompts but rivals get the citation. Fix: the signals in Reranking.

are you quotable?

3 · Synthesize

The model composes one answer from the surviving passages. Clean, claim-first phrasing gets lifted near-verbatim; hedged, meandering prose gets paraphrased away. Fix: Passage Optimization and Extractable Formats.

do you get credit?

4 · Cite

The engine credits a few of the sources it actually used. Which ones (and why) is the closest thing GEO has to a ranking function: Why a Source Gets Cited.

03

Putting It to Work

Diagnose the lost citation, stage by stage

The pipeline's practical use is as a debugging checklist. When a prompt you care about cites someone else, walk the stages in order; the first one you fail is the one to fix.

The citation-autopsy playbook

1

Confirm the engine can see the page at all

Check your server logs or CDN for the engine's crawler and fetch the page as a bot would. Blocked, noindexed, or JavaScript-only content fails before the pipeline even starts. (This is stage 0: AI Crawlers.)

2

Test retrieval: does your passage surface?

Ask the engine the target prompt and close variants; check whether your domain appears anywhere in the sources, even uncited. Also search your key claim verbatim. If you never surface, the problem is chunk quality or semantic match; rewrite the section to answer the question directly.

3

If retrieved but cut: audit the rerank signals

Compare yourself to the pages that did get cited on authority (named author? recognized entity? cited by others?), freshness (when were you last updated?), and answer-fit (do they answer the exact question; do you answer around it?).

4

If cited rarely: check liftability

Read your key passage alone, out of context. If it needs surrounding paragraphs, hedges its claim, or buries the number, the synthesizer can't quote it cleanly. One claim, one passage, answer first.

5

Fix one stage, then re-test on a cadence

Change one variable, wait for a re-crawl, re-run the prompt set. Multi-stage rewrites feel productive but destroy your ability to learn what actually moved the citation.

04

Verification Checks

How to know it's really done

0/3 verified

You own this mental model when missing citations get diagnosed, not guessed at. Click a check to mark it verified:

05

Works Together With

The nodes this one leans on

Subtopics

Tools for this from How AI Engines Work

Go deeper

Always current

These links resolve live: what you get is generated or filtered the moment you click.