Embeddings A list of numbers that stands for what a piece of text means.
Short answer. An embedding is a list of numbers, a vector, that represents the meaning of a piece of text so that texts with similar meaning get similar numbers. Search then becomes arithmetic: embed the question, embed the documents, and return those whose vectors are closest. BGE-M3 produces 1,024 numbers per text and accepts inputs of up to 8,192 tokens in more than 100 languages; nomic-embed-text-v1.5 produces 768 and can be shortened to 64. Embedding an archive is cheap next to generating text, but the index must be rebuilt whenever you change the embedding model, and the vectors themselves can leak information about the text.
01
How it works
An embedding model is a neural network, far smaller than a chat model, trained so that passages with similar meaning end up close together in a space of a few hundred or thousand dimensions. Closeness is measured with cosine similarity or a dot product. The idea began with word vectors: the 2013 word2vec paper showed that useful word representations could be learned in under a day from a 1.6 billion word corpus.
Sentence-BERT, in 2019, made whole-sentence embeddings practical. It cut the time to find the most similar pair among 10,000 sentences from about 65 hours, using BERT to compare every pair, to about 5 seconds, by embedding each sentence once and comparing the vectors.
02
Two open embedding models
Model
Vector size
Longest input
Licence
Worth knowing
BGE-M3
1,024
8,192 tokens
MIT
More than 100 languages; dense, sparse and multi-vector retrieval in one model
nomic-embed-text-v1.5
768, shortenable to 64
8,192 tokens
Apache 2.0
Matryoshka design lets you trade size for a small loss in quality; needs prefixes such as search_document and search_query
03
Storage for a document collection
Each number is normally stored as a 32-bit float, 4 bytes. The raw vectors for a collection are the number of chunks times the vector size times 4 bytes. Indexes add overhead on top.
Chunks
At 1,024 dimensions
At 768 dimensions
At 256 dimensions
10,000
41 MB
31 MB
10 MB
100,000
410 MB
307 MB
102 MB
1,000,000
4.1 GB
3.1 GB
1.0 GB
04
What to check
Use the same model for documents and questions. Vectors from different models cannot be compared.
Rebuild the index when you change the model, or the chunking.
Test the languages you use. BGE-M3 lists more than 100; check Hindi and Hinglish on your own text.
Match chunk length to the model’s longest input, and apply any required prefixes.
Protect the index like the documents. Researchers showed that a multi-step method could recover 92% of 32-token texts exactly from their embeddings, including full names in clinical notes.
05
Why it matters for an office
Embeddings are what let an office search minutes, contracts and policies by meaning. Because embedding models are small, they can run on the same machine as the rest of the system, so documents need not leave the building to be indexed. nomic-embed-text-v1.5 lists about 0.1 billion parameters and the BGE reranker built on BGE-M3 lists 0.6 billion, both far smaller than a chat model.
06
Questions people ask
Is an embedding the same as a token?
No. A token is a piece of text. An embedding is a list of numbers that captures the meaning of a whole passage.
Can an embedding be turned back into the text?
Partly. A 2023 study recovered 92% of 32-token texts exactly from their embeddings with an iterative method. Treat vector indexes as sensitive as the source documents.
Which embedding model should I use?
Pick one that covers your languages and longest chunk, test it on real questions, and keep it fixed. Our model pages list BGE-M3 and nomic-embed-text-v1.5.
Do I need a GPU to create embeddings?
Not necessarily. Embedding models are small, so a CPU can index a modest collection, though a GPU is faster for large archives.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
