What is an embedding?
A customer searches "my card was declined" and your help centre returns nothing, because the article says "payment failed".
Embedding
vector embeddingtext embedding
An embedding is a list of numbers that stands for a piece of text or an image, placed so that similar meanings land close together. OpenAI's docs describe embeddings as vectors of floating point numbers that measure how related two strings are.
Picture a map. Every sentence you own gets a pin, and the pins are placed by meaning rather than by spelling. "My card was declined" and "payment failed" land beside each other. "Card game rules" lands far away, despite sharing a word.
The map has many more directions than two. OpenAI's text-embedding-3-small returns 1,536 numbers per piece of text by default, and text-embedding-3-large returns 3,072. You never read those numbers. The only thing that matters is distance: OpenAI states that small distances suggest high relatedness and large distances suggest low relatedness.
Search stops depending on the exact words
Keyword search matches letters. Ask it for "time off policy" and it will miss the document titled "annual leave". Every support team knows this failure by heart, because customers never use the vocabulary the handbook uses.
Embeddings compare meaning instead. That one change lets people search without knowing your internal terms. It is also the retrieval step under most AI systems that answer from your own files.
- 01Someone finds the right policy without knowing what your company calls it.
- 02Support tickets group themselves by problem, so the top ten complaints surface without anyone reading every ticket.
- 03Recommendations run on similarity, not a rule per product.
- SplitDocuments into passages.
- EmbedEach passage becomes numbers.
- StoreKept for search.
- AskThe question is embedded too.
- RankClosest passages come back.
Change the embedding model and every stored passage has to be redone. Old and new numbers describe different maps.
Related questions
01Do more dimensions make an embedding better?
Not reliably. More numbers cost more to store and search, and OpenAI's larger model is not automatically the right pick for a small collection. Measure retrieval quality on your own questions before paying for the bigger vector.
02Can you get the original text back from an embedding?
No, not directly. An embedding is a lossy summary of meaning, so you store the source text alongside it and use the vector only to find it. Treat the numbers as sensitive anyway. Research has shown text can sometimes be partly rebuilt from them.
03What is the difference between an embedding and a vector database?
An embedding is the list of numbers. A vector database is the place you keep millions of them and search them quickly. You can generate embeddings without a vector database and hold them in Postgres, which is what most projects should start with.
