Key takeaways
- An embedding turns text into a point in a space where distance stands for difference in meaning.
- Semantic search uses embeddings to find matches by meaning, not by exact words.
- Vector databases store embeddings and find the nearest ones quickly.
- Embeddings capture similarity, not facts or named relationships, which a knowledge graph states explicitly.
How embeddings work
Embeddings work by having a neural network turn each piece of text into a long list of numbers, placed so that texts with similar meanings get similar numbers.
Think of a map where every sentence has a location. "We have no budget this quarter" and "There is a budget freeze until spring" land near each other, while "Send me the budget template" lands further away despite sharing a word. The network learns this placement from huge amounts of text, so closeness reflects how language is used.
Semantic search and vector databases
Semantic search finds results by meaning: it turns the query into an embedding, then returns the stored texts whose embeddings sit closest to it.
A vector database is a store built for that job, holding many embeddings and finding the nearest neighbours of a query quickly. The same step drives retrieval-augmented generation, in which the closest passages are handed to a language model so its answer can draw on them.
Embeddings vs knowledge graph
Embeddings capture how similar things are, while a knowledge graph states exactly how things are related.
An embedding can tell you that two call notes are about the same worry; it cannot tell you that one company is a subsidiary of another or that a buyer reports to the finance director. A knowledge graph records those links as named facts that can be queried and checked. Many systems use both: embeddings to find candidates by meaning, explicit relations to keep the facts straight.
An embeddings example from sales calls
An illustrative example is finding every past call note where a buyer raised a security concern, however they worded it.
One buyer asked where data is stored, another mentioned their audit team, a third wanted to know who can see their records. A keyword search for "security" misses most of them. Searching by embedding finds them because the meaning is close, even though the words differ. You still read the results: closeness in meaning is a ranking, not proof that a note is relevant.
Where embeddings fall short
Embeddings are only as good as the model that made them, and they blur distinctions that matter for exact questions.
Two sentences that differ only by a "not" can land close together. Names, numbers and product codes often match poorly by meaning. And the closest text is not always the right one. For questions with one correct answer, such as a price or a contract date, an exact lookup beats a similarity search.
Where Pebbles Ai fits
Common questions
What is a vector embedding?
It is another name for an embedding, stressing that it is a vector, an ordered list of numbers. The two terms are used interchangeably.
Do embeddings store the original text?
No. An embedding is a numeric summary of meaning; systems keep the original text alongside it so they can show or use the passage that matched.
Is semantic search better than keyword search?
It is better at matching meaning and varied wording, while keyword search is better at exact names, codes and phrases. Many search tools combine the two.
About this page. Written by the Pebbles Ai team for people who run sales, marketing and growth. Last updated .
