Embeddings

An embedding model reads a text and returns a fixed-length list of numbers, for example 1,024 of them. No single number means anything on its own; what matters is the position of the whole vector in a space with many dimensions. A sentence about adopting a puppy and one about getting a dog from a shelter land near each other, while a sentence about tax returns lands far away. Distance is usually measured with cosine similarity, which compares the direction of two vectors rather than their length. Think of a map where every text is a pin and neighbors talk about the same thing. Models also exist for images and audio, mapping a photo and its caption to nearby points. A word such as bank can land near rivers or near money, depending on the sentence around it.

This turns fuzzy questions into arithmetic. To search a pile of documents, you embed each passage once and store the vectors; at question time you embed the question, then look for the passages with the nearest vectors. The same idea powers recommendations, duplicate detection, clustering of support tickets and the memory of assistants. It is also the first stage of retrieval-augmented generation, where the passages found are handed to a language model so it can answer from them. Shops, libraries and helpdesks use it for recommendation shelves, semantic search bars and spam or duplicate detection, with a specialized index that finds neighbors quickly instead of checking every stored vector.

Embeddings have limits. A vector keeps meaning but loses exact wording, so a search for a precise code, name or number can miss, and many systems combine vectors with keyword matching. Vectors from different models cannot be compared with each other: change the embedding model and you must embed everything again. Long texts are usually cut into passages first, since one vector for a whole book blurs its content, and the choice of passage size changes what is found. Quality varies between models: some handle many languages while others do best in one, so test with your own material before committing. Typical collections are contracts, support tickets, recipes, song lyrics or source code.

NeverNo’s Vault uses embeddings for your own documents. Each passage, of about 800 tokens, gets a vector of 1,024 numbers computed once by a private model at DeepInfra. The vectors are kept in your browser next to the encrypted documents, and the search itself is done there. GLM 5.3 Flash is a free private model from Z.ai that reads images, and it is one of the models that can then answer from the passages found.

Developers can reach the same embedding model. The API’s /embeddings endpoint returns vectors from qwen3-embedding-0.6b, with 1,024 values and up to 128 inputs a request. The dimensions option shortens the vectors, base64 encoding is understood, and OpenAI’s embedding model names are accepted and served by it. The page about private AI explains where the rest of your data is kept.