What are embeddings?
Embeddings are lists of numbers that an embedding model produces to represent the meaning of text, images or other data, placed so that items with similar meaning sit close together and can be found by vector search.
Checked
An embedding model reads a piece of content and outputs a fixed-length vector, often hundreds or thousands of numbers long. Two passages about the same idea get vectors that point in similar directions, even when they share no words.
That property drives semantic search, retrieval-augmented generation, clustering, deduplication, recommendation and semantic caching. The vector itself means nothing to a person; only distances between vectors carry information.
How embeddings are used
Search and RAG
Documents are embedded once and stored in a vector index. Each query is embedded at request time and the nearest passages are retrieved.
Classification and clustering
Similar tickets, products or contracts group together without hand-written rules.
Caching and deduplication
Near-identical questions or records are detected by distance rather than exact match.
What matters when choosing a model
Fit to the data
General text, code, multilingual content or images. Google's Gemini Embedding 2 preview, released on 2026-03-10, handles text, image, video, audio and PDF in one space.
Dimensions and truncation
Larger vectors cost more to store. Gemini Embedding 2 defaults to 3,072 dimensions and supports Matryoshka-style truncation to shorter vectors.
Switching cost
Changing models normally means re-embedding the whole corpus. The Voyage 4 family, released on 2026-01-15, shares one embedding space across its four sizes.
Open weights
voyage-4-nano is released under Apache 2.0, and open models such as Qwen3-Embedding and EmbeddingGemma 2 can run in a client's own environment.
Rankings move weekly
The MTEB leaderboard on Hugging Face is the common public reference, and its order changes often. A public score is a starting shortlist, not a decision: the model that ranks highest on public benchmarks may not rank highest on a client's contracts or tickets.
Embeddings also have a known blind spot. Exact identifiers, such as part numbers and error codes, are matched poorly by meaning alone, which is why most production systems pair vectors with keyword search.
How Motiontic's engineers apply it
Motiontic's engineers shortlist two or three embedding models, embed a sample of the client's corpus with each, and compare retrieval on an eval set of real questions. Storage, latency, licence and where the model can run are weighed alongside the score.
Embedding model families in use, October 2026
| Provider | Models | Notes |
|---|---|---|
| OpenAI | text-embedding-3-small, text-embedding-3-large | Text embeddings in two sizes |
| gemini-embedding-001, Gemini Embedding 2 (preview) | Embedding 2 is multimodal, 3,072 dimensions by default | |
| Voyage AI (MongoDB) | voyage-4-large, voyage-4, voyage-4-lite, voyage-4-nano | Shared embedding space; nano under Apache 2.0 |
| Open models | Qwen3-Embedding, EmbeddingGemma 2 and others | Self-hostable; compare on the MTEB leaderboard |
Common questions
What is an embedding in simple terms?
An embedding is a list of numbers that stands for the meaning of a piece of content. Content with similar meaning gets similar numbers, so a computer can find related items by measuring distance.
What is the difference between embeddings and a vector database?
Embeddings are the vectors an embedding model produces. A vector database stores those vectors and finds the nearest ones to a query quickly.
Do embeddings need to be regenerated when the model changes?
Usually yes, because vectors from different models are not comparable. Some families, such as Voyage 4, share one embedding space across model sizes within the family.
Which embedding model should a team use?
The one that retrieves well on the team's own data at an acceptable cost. Public leaderboards such as MTEB are a useful shortlist, but rankings change weekly and may not reflect a specific domain.
Related
Vector database
Storage and nearest-neighbour search for embeddings, now in many shapes.
Hybrid search
Keyword ranking and vector similarity, fused into one result list.
Late interaction (ColBERT)
Many vectors per passage, matched token by token: finer retrieval, more storage.
Semantic cache
Reuse an answer when a new question means the same as an old one.
AI product development
New products and major AI features, built from scratch alongside the client's team.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.