Skip to content

Topic 3 min

Retrieval & Embeddings

Your page is chopped into chunks and stored as vectors; the engine pulls the chunk that is the closest semantic match, not the best-ranked page.

Retrieval & …Vector similarityHow chunking changes what is retrievableQuery-to-passage matching

Why it mattersA self-contained, on-topic paragraph can be retrieved even when the page does not rank.

01

Definition & Foundation

What it is, in plain words

When an engine indexes your page, it does not store the page; it chops it into chunks and stores each chunk as an embedding: a list of numbers representing what the text means. At answer time the query becomes a vector too, and the engine pulls the chunks that sit closest in meaning-space. The unit of competition is the chunk, and the match is semantic; a paragraph can be retrieved for a question that shares none of its keywords.

Two practical consequences follow. First, a self-contained paragraph can be retrieved even when its page never ranks; retrieval rewards the best-matching chunk, not the best-ranked URL. Second, structure decides your fate before quality gets a vote: if your answer is smeared across three sections, no single chunk carries it, and no chunk gets retrieved. Citation studies consistently favor self-contained passages of roughly 134–167 words, about the size one chunk holds cleanly.

02

Four Terms, One Picture

How your page looks to a retrieval system

the unit

Chunk

A slice of your page: typically a heading plus the text under it, or a few adjacent paragraphs. Engines retrieve chunks, not pages. If the answer straddles two chunks, neither contains it.

meaning as numbers

Embedding

Each chunk is converted into a vector that encodes its meaning. "How do I stop emails going to spam?" and "improving deliverability" land close together in this space even with zero shared words.

the match

Vector similarity

Retrieval = "find the chunks whose vectors sit nearest the query's vector." Nearness is meaning-likeness. This is why answering the actual question beats repeating the keyword.

the mindset shift

Semantic match vs keyword match

Classic SEO asked "does the page contain the words?" Retrieval asks "does this chunk mean the answer?" Synonyms, paraphrases, and plain language all match; keyword stuffing adds nothing.

03

Putting It to Work

Structure pages so the right chunk exists

You can't control how an engine chunks your page, but you can write so that any reasonable chunking yields a self-contained, retrievable answer. That is a structural craft:

The chunk-friendly structure playbook

1

One question per section

Give every section a single job, stated in its heading (ideally as the question users ask). A section that answers three questions poorly beats none of the three chunks that would each answer one well.

2

Answer within the first two sentences under the heading

Chunk boundaries often fall at headings, so the heading plus its opening lines is the most reliably intact unit. Put the direct answer there; elaborate below.

3

Make each passage survive amputation

Resolve pronouns, repeat the entity name, restate the subject; the chunk may be read with zero surrounding context. "It also supports this" is dead weight; "Acme also supports SAML single sign-on" retrieves.

4

Keep key passages in the 130–170 word zone

Long enough to fully answer, short enough to fit a chunk cleanly. If a key answer runs 400 words, extract its essence into a lead paragraph and let the detail follow.

5

Write for the paraphrases too

Semantic match means "email warm-up," "sender reputation," and "why do my emails bounce" can all pull the same chunk, if the chunk speaks plainly. Use the words real users use, not only your product's vocabulary.

04

Verification Checks

How to know it's really done

0/3 verified

Your structure is retrieval-ready when chunks survive on their own. Click a check to mark it verified:

05

Works Together With

The nodes this one leans on

Tools for this from How AI Engines Work

Go deeper from The RAG Pipeline

Always current

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

Sources