Softanza
StartFrançais
Softanza
StartFrançais
Reference · Language & intelligence

Neural tier

20 functions in 7 classes, grouped by what they do, each with the explanation the library gives of itself.

The story of this area, for a general reader: what a crafter does with it, an example run and its rated lanes, is on its Platform page.

The functions, by what they do

Testing and checking

  • IsConstrainingDecoding Will the NEXT live call be grammar-constrained? Ask this rather than assuming: a nested structure, an off switch, or a build without the sampler rung all answer 0, and WhyNotConstrained() says which. in stzLLMFunction
  • Check The family's rule face: run me, hand me back the findings. in stzOutputSchema
  • IsDecodingConstrained Does anything actually CONSTRAIN DECODING with this grammar? Ask before reporting an answer as grammar-constrained. It answered 0 for as long as that was true; gbnf_machine.zig is the rung underneath it, so it answers 1 since 2026-08-20 -- the ENGINE is the authority on that number, never this comment. in stzOutputSchema
  • IsExpressibleAsGrammar TRUE when this declaration can be expressed as a grammar at all. in stzOutputSchema
  • IsLoaded load / unload (runtime gguf) in stzNeuralModel
  • IsGenerative TRUE if this model can GENERATE text (a causal decoder). in stzNeuralModel
  • IsReady stateless facade over the ggml inference backend. TRUE if the ggml backend is compiled in and executes. in stzNeuralEngine

Finding

  • LastFindings The findings the LAST validation produced, in the family's unified shape -- so a refusal can be handed to stzRuleReport.Ingest() and stand in the same CI gate as every other rule in the library. in stzLLMFunction
  • SearchByVectorXT The same search from a query embedding the caller already holds (the companion of AddEmbedded). ONE search path serves both doors: the route -- CPU resident dataset or GPU resident corpus -- is decided here and nowhere else. in stzSemanticIndex

Counting and measuring

  • Count feeding the index in stzSemanticIndex

Replacing and setting

  • SetRetrySampling The sampling a RETRY uses. Attempt 1 never sees these -- it is greedy. in stzLLMFunction

Adding

  • AddGolden golden sets in stzLLMFunction
  • AddText Embeds at ADD time (the model must be loaded), so a bad text fails loudly here and not in the middle of a search. in stzSemanticIndex
  • AddTexts feeding the index in stzSemanticIndex
  • AddEmbedded Index a text with an embedding the CALLER already has -- another model, a cache, a wire -- so the index never insists on embedding it again. The vector must be L2-normalised (the score formula assumes it) and of the index's dimension. in stzSemanticIndex

Converting

  • ToGBNF The GBNF text for this declaration, or a raise naming the construct that cannot be expressed. Nested structures are REFUSED rather than flattened: a grammar that accepted what this schema rejects would put the two layers into disagreement. in stzOutputSchema
  • ParseOutput RAW MODEL TEXT to a validated whole, or to a refusal that names the rule it failed. [ :ok, :value, :findings, :shape ] in stzOutputSchema

Reading and writing

  • LoadFrom (Named LoadFrom, not Load -- "Load" collides with Haro's load keyword.) in stzNeuralModel
  • LoadStatus Why the last LoadFrom was refused (0 ok, -1 unreadable, -2 digest mismatch, -3 no recorded digest). in stzNeuralModel

Walking and applying

  • RunGoldens Goldens hold STRUCTURES as readily as scalars now. Two things had to change for that, and both were real defects rather than gaps: Haro's own `=` answers 0 for two identical lists, so a structured golden could never have passed; and a failing structured case that reports only "expected / got" is unreadable, so the failure now carries the FIELD that moved. in stzLLMFunction

The explanations are the ones the library gives of itself, read from its documentation comments on 2026-10-01 04:14, from the library at commit 0e72e2e2c, by the library's own stzSelfDoc over 635 class sources.

The classes of this area

Narrations about this area