What are embeddings?

Embeddings: definition

A computer does not understand words by itself, but it does understand numbers. An embedding model therefore converts a text into a vector, a list of several hundred to several thousand numbers. These numbers encode what the content is about, not which letters appear in it.

You can think of it like coordinates on a map of meanings: “heating pump defective” and “circulation pump not running” sit close together, “vacation request” far away. Embeddings exist for text, images, audio and code. They are a model type of their own, not a language model that writes text.

How embeddings are used

All content, such as manuals, product descriptions or emails, is converted into vectors once and stored. For a search query, the question is also turned into a vector. The system then finds the content whose vectors are most similar to that of the question; cosine similarity is commonly used for this.

Typical applications are semantic search, recommendations, detecting duplicates, grouping inquiries and above all retrieval-augmented generation (RAG): the most relevant passages are found for the question and given to a language model as the basis for its answer. The advantage: the answer is based on your own documents, and the sources used can be cited.

Real-world example: searching a service manual

A machinery manufacturer maintains manuals, service reports and damage reports for its equipment. Keyword search fails when technicians type “machine hums on startup” but the document says “bearing noise at startup.” With embeddings, the search finds both, because the two wordings are close in meaning.

In the project, the documents were split into sensible sections and a vector was calculated for each section. Classic keyword search stayed active as well, because it finds article numbers and error codes more reliably. The combination of both is called hybrid search. Which mix works best is decided by technicians’ real search queries, not by theory.

Limits and what to watch for

Vectors from different models cannot be compared: if you change the embedding model, all content has to be recalculated. Technical terms and multilingual support also depend on the model. And during processing your texts reach the embedding model, so check data protection, data processing agreements and server location.

Nano Concepts selects models and splits content to suit your material and tests the quality of the results with real search queries from your daily work. This includes splitting documents into sensible sections and reviewing results regularly so that search stays reliable as new content arrives.

• Split documents into sensible sections rather than page blocks
• Choose an embedding model that fits your language and field
• Keep keyword search for numbers and codes
• Test result quality with real questions from daily work
• Recalculate all vectors when you change models