Connecting AI to your data · Layer 3

What is an embedding?

Embeddings (vector representations)

A list of numbers that captures the meaning of a text: two texts with similar meanings get similar numbers.

A way to picture itImagine a vast map where every sentence is placed according to its meaning: “contract termination” sits next to “ending the agreement”, with no word in common. Searching means looking at what lies near your question.

The diagram · Embeddings

01Split documentsChunks of a few paragraphs

02Embedding modelTurns each chunk into a vector

CheckSame protection as the source

03Vector databaseVectors stored with their source document

04Question as vectorConverted into numbers the same way

05Nearest neighboursThe chunks closest in meaning

How it works

01A specialised model reads each chunk of text and turns it into a list of several hundred or several thousand numbers, a vector, that sums up its meaning.
02These vectors are stored in a vector database. The user’s question is converted in the same way, and the chunks whose vectors are closest are retrieved.
03Search therefore works on meaning rather than keywords. It is the search engine behind RAG, often combined with conventional keyword search, which is more reliable for proper names and references.

In a business function

Customer service

An insurer indexes its 3,000 procedure sheets using embeddings. An adviser types “the customer cracked their phone screen” and finds the sheet on “accidental damage to portable devices”, which a keyword search would have missed. Searching for a contract number, on the other hand, fails: codes are better found by keyword. The tool finally chosen therefore combines both.

The compliance angle

A vector computed from personal text is still personal data: researchers recovered 92% of 32-token texts word for word from their embeddings alone, and retrieved full names from clinical notes (United States · Cornell University, EMNLP 2023). The vector database must therefore be protected like the original documents: encryption, access control, retention period (EU · GDPR, art. 32). A deletion at source must carry through to the index, otherwise the erased data can still be queried. Vector and embedding weaknesses rank among the ten main risks for large language model applications (international · OWASP).

Limits to know

Meaning-based search is less reliable for codes, references and proper names.
Switching embedding model means recomputing the whole index.
Closeness in meaning is not relevance: a neighbouring chunk may be off-topic or out of date.

The question to ask your supplier

“Where are the embedding model and the vector database hosted, how are our documents’ access rights applied there, and how quickly is a deletion at source reflected?”

Submit a brief

Where this concept sits

01Business questions
02Definitions and owners
03Connected data
04AI
05Rights and human sign-off
06Evidence

Related concepts

Frequently asked questions

Is the embedding model the same as the LLM?

Usually not: it is a smaller, specialised model, often from another vendor. It can run on your own servers even if the LLM is hosted elsewhere.

Can the text be recovered from a vector?

Partly, yes: published research shows that a vector can be turned back into text with good fidelity. It should therefore be treated as sensitive data, not as an anonymous fingerprint.

Is it the same thing as a semantic layer?

No. Embeddings bring texts together by statistical resemblance; a semantic layer sets explicit business definitions, each with an owner. The two complement each other.

Educational explainer, not legal advice. Up to date as of 29 October 2026.

What is an embedding?

Embeddings (vector representations)

A list of numbers that captures the meaning of a text: two texts with similar meanings get similar numbers.

How it works

01A specialised model reads each chunk of text and turns it into a list of several hundred or several thousand numbers, a vector, that sums up its meaning.

02These vectors are stored in a vector database. The user’s question is converted in the same way, and the chunks whose vectors are closest are retrieved.

03Search therefore works on meaning rather than keywords. It is the search engine behind RAG, often combined with conventional keyword search, which is more reliable for proper names and references.

In a business function

Customer service

The compliance angle

Frequently asked questions

Is the embedding model the same as the LLM?

Usually not: it is a smaller, specialised model, often from another vendor. It can run on your own servers even if the LLM is hosted elsewhere.

Can the text be recovered from a vector?

Partly, yes: published research shows that a vector can be turned back into text with good fidelity. It should therefore be treated as sensitive data, not as an anonymous fingerprint.

Is it the same thing as a semantic layer?

No. Embeddings bring texts together by statistical resemblance; a semantic layer sets explicit business definitions, each with an owner. The two complement each other.