Definition
Vector search is the lookup that finds the closest matches in a set of numeric representations, returning nearest neighbours rather than exact hits. Semantic search names the effect a user experiences; vector search names the operation performed. It runs under every AI retrieval step you care about.
At a glance
- CategoriesTechnical
- Related fieldsAI Search, Technical
- DifficultyAdvanced
At a glance
- Vector search compares numbers, not text, and has no notion of what a document says.
- Results rank by distance, so there is always a nearest match and never an empty one.
- Indexes are approximate by design, trading accuracy for speed at scale.
Nearest neighbours in a number space
Every chunk of text becomes an embedding and gets stored. A question goes through the same conversion, and the search becomes a geometry problem: which stored points sit closest. From the user’s side that is semantic search.
Comparing against every point would be too slow at web scale, so the index groups similar points and searches only promising regions, which is what libraries like Faiss exist to do. OpenAI’s embeddings guide describes the setup: store document vectors once, compare the query vector at search time.
Approximate is the operative word: the index usually finds the true nearest matches and occasionally misses one.
Close enough wins, and nothing returns empty
Keyword search tells you when it has nothing. Vector search never does, and that changes what failure looks like.
A query always has a nearest neighbour, even when that neighbour is barely related. The system fails by returning three mediocre passages with the confidence it would show for three excellent ones, and the model writes from them anyway.
So adequacy is not enough. There is no bar to clear. You are trying to be closer than everyone else writing about the same thing, and distance comes down to how tightly your chunk sits on one subject.
It is also why a reranking stage exists: distance is a rough proxy for usefulness, so something slower reads the finalists.
Land in a tight neighbourhood
You never touch a vector. You control the text that becomes one.
- Keep a section to one subject. Cover three and it sits in the average of three positions, close to nothing.
- Put the subject in the first line. Opening sentences carry disproportionate weight in where a chunk lands.
- Use your buyers’ vocabulary, not the internal one. A question in customer language lands near text in customer language.
- Give every section a heading. Headings often mark where one chunk stops and the next starts.
Where vector search ends and semantic search begins
| Vector search | Semantic search | |
|---|---|---|
| What the term names | The lookup operation | The user-facing behavior |
| What it works on | Numbers in an index | Meaning in your content |
| Who uses the word | Engineers | Marketers |
| What you change | Nothing directly | Clarity of each section |
| Can it run alone | Yes, on any vectors | Only via a lookup like this |
Frequently asked questions
-
Do I need a vector database for my website?
Only if you are building something that searches your own content. It does nothing for your visibility in ChatGPT or Google, which build their own vectors from your pages and never look at your infrastructure.
-
Why does a vector search never come back empty?
Because it ranks by distance rather than testing a condition. There is always a closest point, however far away. Some systems add a cutoff, but the operation itself cannot return nothing, which is why a weak match and a strong one look alike.
-
Can I see how close my pages are to a given question?
You can approximate it with an embeddings API: embed the question, embed your sections, compare distances. An afternoon of work, useful for spotting a section sitting nowhere near the question it was written to answer. It will not reproduce what an assistant does.
-
Is understanding vector search worth a marketer's time?
The shape of it, yes. Knowing that retrieval compares chunks by distance, always returns something and rewards specificity changes how you structure a page. Past that it stops paying.