Struct TunedTagger 

pub struct TunedTagger { /* private fields */ }

A tuned tagger ready for inference: weights + the label spaces they were trained with.

Implementations

impl TunedTagger

pub fn load( dir: &Path, base_dir: &Path, tokenizer: &Path, max_len: usize, ) -> Result<TunedTagger, String>

Load from a save() directory (tagger.safetensors + tagger.json) plus the base config and a tokenizer. The label space comes from the checkpoint, never from a hardcoded list, so a model trained on a different corpus’s spec decodes correctly.

pub fn labels(&self) -> &[String]

pub fn enable_relations( &mut self, dir: &Path, spec: &VocabularySpace, ) -> Result<(), String>

Attach the trained biaffine relation head (Head C) so projection binds arguments and emits dimension-2 tokens. dir is the checkpoint directory holding relations.safetensors.

pub fn has_relations(&self) -> bool

pub fn project(&self, sentence: &str) -> Result<Situation, String>

Project a sentence into a [Situation] for the roaring index — the bridge from the tuned tagger to ingest. Emits the six Vocabulary-Space dimensions this tagger can produce:

• dim 1 typed entity tokens (org/raytheon) from Head A’s facets
• dim 3 geo/… / time/… via the deterministic loci normalisers
• dim 4 qty/… bucket URIs via the units canonicaliser
• dim 5 state/negated / state/hedged from a STATE cue span

Negation scope: an epistemic cue scopes its clause, so a state/negated cue sets negative infon polarity on the other tokens of the sentence, not merely on the cue. That is what makes Bel/Pl meaningful — “Raytheon does not manufacture Aegis” must not assert org/raytheon positively about that relation.

pub fn tag(&self, text: &str) -> Result<Vec<PredictedSpan>, String>

Tag a sentence → typed spans with epistemic polarity. BIO decoding merges B-X + following I-X; each span’s epistemic class is the majority vote of its tokens.