How does experiential-output achieve structured output? What is the principle behind it

[little077]

How does experiential-output achieve structured output? What is the principle behind it

[ahuaiqom]

Hi @little077 👋

“Structured output” is achieved by turning your target schema (e.g. JSON Schema) into a grammar and then doing constrained decoding so the model can only emit tokens that keep the output valid. If the partial output ever deviates, the decoder blocks those tokens and nudges the model back onto a valid path; optionally a validator/repair step runs at the end.

How it works (principle)

1.Schema → Grammar

• Your spec (JSON Schema / Pydantic / TypeBox, etc.) is compiled into a grammar/finite-state machine (think EBNF).

• The grammar defines exactly which characters/tokens are legal at every position.

2. Constrained decoding

• At generation time, the sampler masks out tokens that would violate the grammar.

• The model is still “free” within the allowed set, but can’t leave the schema (e.g., must open/close braces, produce valid numbers/enums, required fields, etc.).

3. Validation & repair (optional)

• The produced string is validated against the schema.

• If invalid or incomplete, a lightweight self-correction pass (or a guided retry with the same grammar) fixes it.

4. (Optional) Tool/field hints

• System instructions or field descriptions are surfaced to the model to improve content quality, while the grammar guarantees the shape.