Development build. This page describes main, not a released package. The latest published Lodestar.Embeddings is 0.7.0 — read its documentation.

Home › Embeddings › Tokenization

TruncationStrategy

Which end is cut when a sequence is longer than MaxLength.

public enum TruncationStrategy { None, LongestFirst, Right }

Members — None does not truncate, so a sequence over the cap stays over it and the model decides what to do. LongestFirst cuts the longest sequence of a pair first, which is what a sentence-pair model wants so that neither half disappears. Right cuts from the end, keeping the beginning.

Example — the default cap with the end cut.

using Lodestar.Embeddings.Tokenization;

var options = new EncodingOptions
{
    MaxLength = 8,
    Truncation = TruncationStrategy.Right,
};

TruncationStrategy strategy = options.Truncation;  // => Right

Remarks — Right is right for most things, because the start of a document usually carries more of what identifies it. It is not universal: for a question-and-context pair, cutting the end removes the context and keeps the question, which is why LongestFirst exists.

None is the honest default for a caller who would rather see a failure than a silently shortened input — a truncated document embeds to a vector for a different text, and nothing downstream can tell.

Applies to — net10.0, netstandard2.0.

See also — EncodingOptions, BatchEncoder.