An embedding is a dense vector representation of data — text, image, audio, structured entity — in a learned continuous space where semantic similarity maps to geometric proximity. Two sentences with similar meaning get vectors close to each other; two unrelated sentences get vectors far apart. The geometry is the entire point: search, clustering, classification, recommendation, anomaly detection all become tractable once you have a good embedding space, because the hard work of representing meaning has been pushed into the encoder and what remains is linear-algebra geometry over fixed-dimension vectors.
The embedding landscape moves fast. **Text embeddings** are dominated by vendor APIs (OpenAI text-embedding-3-small/large, Cohere embed-v3 / embed-v4, Voyage, Jina, Google Gecko) and strong open-weight alternatives (BGE, GTE, Nomic, E5, Jina open). **Multimodal embeddings** (CLIP, SigLIP, Jina CLIP) map images and text into a shared space. **Code embeddings** (Voyage code, OpenAI embedding-3 with code specialisation, Codestral Embed) specialise for source-code similarity. Dimensions range from 384 to 3072; larger is not always better, and many 2024+ embedders support truncation (Matryoshka embeddings) that let you trade dimension for retrieval cost at query time.
For APAC mid-market teams, the practical embedding decisions are: which vendor or open model, which dimension, and whether to fine-tune. **Vendor APIs** are the right starting point — they give state-of-the-art quality with no MLOps cost. **Open models** (served via Hugging Face TEI or self-hosted) make sense when data residency requires local processing, when throughput is very high, or when the cost curve flips. **Fine-tuning** a domain embedder (on pairs of query-relevant passage from your corpus) is a real quality lever but requires labelled training data and evaluation infrastructure — often worth it for technical, legal, or medical corpora where generic embedders underperform.
The non-obvious operational trap: **embedding versions are not interchangeable**. A corpus embedded with OpenAI text-embedding-ada-002 cannot be searched with text-embedding-3-large — the geometry is different, and you must re-embed the whole corpus to upgrade. Plan for this cost up front, and keep the embedding model version in metadata alongside every stored vector. Multi-version corpora (some vectors old, some new) produce silently-wrong retrieval results.
Where AIMenta applies this
Service lines where this concept becomes a deliverable for clients.
Beyond this term
Where this concept ships in practice.
Encyclopedia entries name the moving parts. The links below show where AIMenta turns these concepts into engagements — across service pillars, industry verticals, and Asian markets.
Other service pillars
By industry