ITokenEmbeddingModel<T>

Interfaces · AiDotNet.Interfaces

Defines the contract for a model that produces per-token (contextual) embeddings for a text.

Known implementations: EmbeddingModelTokenAdapter<T>

For Beginners

Think of this as describing a sentence word-by-word instead of summarising the whole sentence in one number-list.

  • Single-vector model: “The quick brown fox” -> one vector.
  • Token embedding model: “The quick brown fox” -> four vectors (one for each word).

Keeping a vector per word lets a reranker ask, for every word in your question, “which word in this document matches me best?” - a much finer-grained comparison than a single overall vector.

How It Works

Unlike a single-vector embedding model (IEmbeddingModel) which collapses an entire text into one dense vector, a token embedding model returns one vector per token. This is the representation required by late-interaction retrievers and rerankers such as ColBERT (Khattab & Zaharia 2020), where the relevance of a (query, document) pair is scored by matching individual query tokens against individual document tokens rather than comparing two pooled vectors.

Properties

Property Summary
int EmbeddingDimension { get; } Gets the dimensionality of each per-token embedding vector.

Methods

Method Summary
Vector<T>[] EmbedTokens(string text) Produces one embedding vector for each token in the supplied text.