Vector Search

A classic search compares sequences of letters. Anyone who types “fix a bicycle” finds pages that contain exactly these words. A text about “patching tires on a bike” gets overlooked in the process, even though it fits perfectly. Vector search solves this problem differently. It has each piece of content translated beforehand by an AI model into a long list of numbers that describes its meaning. Such lists of numbers are called vectors. Two pieces of content with similar meaning receive similar numbers, and the search delivers everything that comes close to the query.

Why searches fail at the level of words

Language is messy. There are many words for the same thing, and the same word can mean different things. A pure word search must therefore be retrofitted with synonym lists and rules. Someone maintains these lists by hand, and they are never complete.

Vector search doesn’t need such lists. It has learned from huge amounts of text which phrasings appear in similar contexts. That’s why it also works across language boundaries. A German question can find an English paragraph if the model has learned both languages.

This became especially important with chatbots. A language model knows nothing about a company’s internal documents. But it can be given matching text passages before answering. Retrieving these passages is almost always handled by a vector search. This approach is called Retrieval Augmented Generation, or RAG for short, and is today the standard way to connect a language model with proprietary knowledge.

From meaning to a series of numbers

The first step is called embedding. A specialized model reads a text passage and outputs a vector, often with a few hundred to a few thousand numbers. You can picture this as a point in a very high-dimensional space. Texts about bicycles end up in one corner, texts about recipes in another.

When a query comes in, the question is likewise turned into a vector. Then the system searches for the points with the smallest distance. Usually, one doesn’t measure the straight-line distance for this, but rather the angle between the series of numbers. Only the principle matters: small distance means similar meaning.

With millions of entries, it would be too slow to compare each one individually. That’s why vector databases create a kind of map over the points, an index. The search first jumps to the appropriate area and only compares precisely there. This is called an approximate search, because it occasionally misses a good match. In exchange, it responds in milliseconds instead of minutes. A common misconception is that vector search completely replaces word search. In practice, both are combined, because names, item numbers, and dates need to be found exactly.

Vector search in products and headlines

You encounter it constantly in everyday life without hearing its name. The photo gallery on your phone finds pictures with dogs, even though nobody labeled them. Streaming services suggest similar songs. Shop searches understand “warm jacket for autumn”, even if the product title says “quilted coat”.

In companies, the technology is embedded in assistants for manuals, contracts, and support tickets. An employee asks in plain language, the system searches for the matching paragraphs, and a language model formulates the answer. This exact setup explains why providers of vector databases like Pinecone, Weaviate, or Qdrant show up in business news. Even established databases like PostgreSQL or Elasticsearch have retrofitted vector capabilities.

For you as a reader of tech news, this means: when semantic search, embeddings, or RAG are mentioned, it’s essentially about vector search. And when a chatbot cites the wrong source, the error often doesn’t lie in the language model. Frequently, the search delivered the wrong text passages beforehand.