TokenFsmConstraint

Models & Types · AiDotNet.Serving.StructuredOutput

An ITokenConstraint backed by a compiled deterministic finite automaton (DFA) over token ids. Higher-level compilers (regex, JSON-schema, grammar) lower their language to a token-level DFA and hand it to this class, which does the per-step masking and state advance.

Implements: ITokenConstraint

For Beginners

think of this as a board game where the current square lists which moves (tokens) are legal. The model picks among the legal moves; illegal ones are hidden. When you land on a “finish” square you’re allowed to stop.

How It Works

Each state has a set of outgoing transitions tokenId -> nextState. In an accepting state the end-of-sequence token is additionally permitted, which is how the engine learns it may stop. The DFA is tokenizer-agnostic: it speaks only token ids, so it composes with any tokenizer the serving layer registers.

Constructors

Constructor Summary
TokenFsmConstraint(IReadOnlyList<IReadOnlyDictionary<int, int>> transitions, IEnumerable<int> acceptingStates, int eosTokenId, int startState) Creates a token-level DFA constraint.

Properties

Property Summary
bool IsComplete { get; }
bool IsTerminal { get; }

Methods

Method Summary
TokenFsmConstraint FromChoices(IEnumerable<IReadOnlyList<int>> choices, int eosTokenId) Builds a constraint that restricts the output to exactly one of a fixed set of token sequences (a choice / enum constraint).
TokenFsmConstraint FromSequence(IReadOnlyList<int> tokenIds, int eosTokenId) Builds a constraint that forces an exact token sequence, then end-of-sequence.
void Accept(int tokenId)
void ApplyMask(Span<float> logits)