What are embeddings, and why does almost every AI feature depend on them?
Embeddings are the quiet infrastructure underneath semantic search, recommendations, clustering and retrieval. Understanding them is the fastest way to reason about why an AI feature returns what it returns.
By Quality AboveAll · · 7 min read
- An embedding is a list of numbers that positions a piece of content in space so that similar meanings sit close together.
- Similarity is measured by distance, which is why search can match a question to an answer that shares no words with it.
- Embedding choice, chunk size and normalisation affect retrieval quality more than the language model sitting downstream.
The core idea
An embedding model reads a piece of content and outputs a fixed-length list of numbers, often several hundred or a couple of thousand of them. Those numbers are coordinates. The model is trained so that content with similar meaning lands in a similar place, which means distance between two embeddings becomes a usable measure of relatedness.
This is what allows a search for "how do I get my money back" to return a document titled "Refund policy" that contains none of those words. The two sit near each other in the space because the model learned that they mean roughly the same thing.
Why this beats keyword matching, and where it does not
Keyword search fails on synonyms, paraphrases and questions phrased differently from the source, which is most of how humans actually ask things. Semantic search handles all three naturally, and that alone makes it worth adopting for any body of text people query in their own words.
It is weaker exactly where keyword search is strong: exact identifiers. Product SKUs, error codes, invoice numbers and surnames are things you want matched literally, and a semantic model will happily return something that merely feels similar. This is why serious retrieval systems run both and merge the results, as covered in our guide to RAG.
Semantic search understands what you meant. Keyword search knows what you typed. Production systems need both, because users do both.
The decisions that change quality
Chunk size is the first and most consequential. Embed an entire document and its meaning averages into vagueness; embed a single sentence and you lose the context that made it interpretable. Chunks that follow the document's own structure, a section, a policy clause, a step, generally outperform fixed character counts.
Model choice is the second. Larger embedding models capture nuance better but cost more per item and produce larger vectors to store and search. Open model catalogues make it cheap to test several against your own content, which is the only benchmark that matters, because published leaderboards were not built on your documents.
What embeddings are used for beyond search
Deduplication and clustering fall out of the same property: near-identical support tickets sit near each other, so you can group them without writing rules. Classification works similarly, by comparing a new item to labelled examples rather than training a bespoke classifier.
Recommendation systems use the same trick over products or content, which we cover in building a recommendation engine. In each case the value comes from the same place: meaning became arithmetic, and arithmetic is cheap.
Practical cost and storage notes
Embedding is a per-item cost you pay on ingestion and again on every re-index, so the size of your corpus and how often it changes drives the bill more than query volume does. Caching embeddings for unchanged content is the obvious saving that teams routinely forget to implement.
Storage scales with dimensions multiplied by item count, and higher dimensions also slow search. Many teams find a mid-sized model with dimension reduction gives almost all the quality at a fraction of the footprint, which matters once a corpus is measured in millions of chunks. Where to keep them is covered in vector databases explained.
Frequently asked questions
What is the difference between an embedding and a vector?
In this context they are used interchangeably. The vector is the list of numbers; the embedding is that vector when it represents the meaning of some content.
Can I compare embeddings from two different models?
No. Each model organises meaning differently, so distances between vectors from different models are meaningless. Switching models requires re-embedding the whole corpus.
Do embeddings work for languages other than English?
Multilingual embedding models handle many languages and can even match a question in one language to an answer in another. Verify quality on your own content, as performance varies considerably by language.
Getting weak results from a semantic search feature? A free technical review will pinpoint whether the problem is chunking, the embedding model, or what happens after retrieval.
