Search a keyword-based system for "how to stop feeling anxious before a talk" and it will look for pages containing those specific words. Miss the exact phrasing — the source you actually need says "manage nerves before public speaking" — and a keyword match can walk right past it, even though it's clearly the answer. A vector index exists to fix exactly this problem: it lets a system match on what something means, not on which words happen to appear.
The old way: matching words
Traditional search works by matching text. It builds something like an index at the back of a book — this word appears on these pages — and when you search, it looks for documents containing your search terms, weighted by how often and how prominently those words appear. This works reasonably well when you know the exact vocabulary the content uses. It works poorly whenever there's a mismatch between how you asked and how the answer is phrased — synonyms, rewording, a different technical term for the same concept, a question phrased conversationally against content written formally.
The new way: matching meaning
A vector index takes a different approach entirely. Instead of storing text as a list of words, it converts each passage into a long list of numbers — called a vector, or an embedding — that represents what that passage means. This conversion is done by a model trained specifically so that passages with similar meaning end up with similar numbers, even if the actual words are completely different. "Reduce inflammation" and "lower swelling" end up close together in this numerical space. "Bank" the riverbank and "bank" the financial institution end up far apart, because the model has learned they mean different things despite sharing a word.
When you search, your question goes through the same conversion, becoming its own vector. The system then looks for the stored passages whose vectors are mathematically closest to your question's vector — closeness in this numerical space corresponds to closeness in meaning. That's what people mean by "semantic search" or "searching by meaning": the match happens on concepts, not on shared vocabulary.
Why "meaning" is the right word and not marketing gloss
It's worth being precise about what's actually happening, without overselling it: the model isn't "understanding" the way a person does. It has learned, from a vast amount of text, statistical patterns about which words and phrases tend to occur in similar contexts — and that turns out to be a remarkably good proxy for meaning in practice. Two passages that a human would agree are "about the same thing" reliably end up close together in this numerical space, and that's the property that makes it useful for search, whatever you want to call the mechanism underneath it.
Why this matters for AI-generated answers specifically
A vector index isn't just a nicer search experience for a human typing into a box — it's the retrieval half of how an AI system answers a question at all. When a language model needs to ground its answer in real, current information rather than guessing from what it memorised during training, it needs a way to pull the right passages out of a huge stored collection, fast, based on what the question is actually asking rather than the specific words used to ask it. A vector index is what makes that retrieval step possible. See how AI search finds an answer for how retrieval by meaning connects into an actual generated answer, and what is retrieval-augmented generation for the technique this whole approach is built on.
Where this fits without giving away the machinery
Exactly how a system chooses to represent meaning, how it organises billions of these vectors for fast lookup, and how it balances precision against speed are genuine engineering problems with a lot of nuance to them — and not the part that matters to a reader trying to understand what's happening when they search. What matters is the shape of the idea: content gets stored by what it means, a question gets converted the same way, and the system finds what's closest. askFinz's web index is built around exactly this principle — every page, document, and passage it reads is stored so it can be found by meaning, which is a large part of why a plainly worded, conversational question tends to find the right answer even when it doesn't share a single word with the source that actually contains it.
Further reading
- How retrieval by meaning becomes an actual written answer: How AI search finds an answer
- The technique that grounds a model's answer in retrieved sources: What is retrieval-augmented generation (RAG)?
- The full pipeline a page goes through before it's stored this way: How a search index is built