What are embeddings and vector databases?

What an embedding is

An embedding model converts a text section into a list of numbers – often several hundred to a few thousand values. Each value is not meaningful on its own; what matters is the position of the vector in this high-dimensional space. Texts with similar meaning end up close to each other, even if they use different words. "Holiday request" and "apply for leave" are therefore found together, although they share no word.

How semantic search works

1. Split: documents are divided into sections (chunks), for example paragraphs or pages.
2. Embed: each section is converted into a vector with an embedding model.
3. Store: the vectors are saved together with metadata such as source, date and permissions in a vector index.
4. Search: a question is converted into a vector as well; the database returns the sections with the smallest distance.
5. Use: the hits are passed to a language model, which formulates the answer – the principle behind RAG.

Vector database or vector index?

The term vector database is used broadly. It can mean a specialised standalone system, but also a vector index within an existing database. What matters in practice is less the product category than a few properties: fast nearest-neighbour search, filtering by metadata and permissions, and the ability to update or delete entries reliably.

Why this matters for companies

Classic keyword search fails when employees phrase things differently from the document. Semantic search finds content by meaning and, depending on the model, across language boundaries. At the same time, vector search alone is not always enough: exact terms such as article numbers or names are often found better with keyword search, which is why many systems combine both.

Limits and points to check

• Chunking: sections that are too large dilute the meaning, sections that are too small lose context.
• Model consistency: query and documents must be embedded with the same model; switching models requires re-indexing.
• Permissions: the search must filter by the rights of the person asking, otherwise confidential content appears in answers.
• Deletion: removing a document must also remove its vectors.
• Data protection: embeddings are derived from the original text and should be treated with the same care – see GDPR-compliant AI.

Embeddings in AIPE

In AIPE, documents are extracted, split into sections and embedded as vectors with 1536 dimensions. Vector search is filtered by the user's permissions (role-based access control), so only documents the person is allowed to read are considered. Answers reference the source documents. In addition, an ontology links customers, projects, products and processes into a knowledge graph. More in the guide AI knowledge base and on the knowledge page.