Struct IntentClassifier 

pub struct IntentClassifier { /* private fields */ }

Intent classifier using an ONNX model (all-MiniLM-L6-v2 or DistilBERT).

Implementations

impl IntentClassifier

pub fn load( model_dir: &Path, allow_unverified: bool, trust_mode: TrustMode, ) -> Result<Self, ClassifierError>

Load classifier from model directory.

Expected directory structure:

model_dir/
├── intent_classifier.onnx
├── tokenizer.json
├── config.json
├── calibration.json or temperature.json (optional)
├── checksums.json
└── version.txt

Arguments

• model_dir — Resolved model directory (output of NL02 resolver chain).
• allow_unverified — Operator escape hatch. When false (NL04 default per FR-7), missingness is fatal. When true, missingness downgrades to tracing::warn!. Tampering on a present file ALWAYS errors regardless of this flag — see the inline contract documented at [verify_integrity].
• trust_mode — Output of [TrustMode::from(ResolverLevel)]. Trusted mode anchors integrity in the binary’s baked-in manifest; Custom mode trusts the user-supplied manifest.json shipped alongside the model directory.

Errors

Returns ClassifierError if:

• Model files not found
• Checksum verification fails (AC-11.8 / NL04 integrity contract)
• ONNX Runtime initialization fails

pub fn classify( &mut self, input: &str, ) -> Result<ClassificationResult, ClassifierError>

Classify intent from natural language input.

Critical: batch_size=1 enforcement (C1 mitigation)

ONNX Runtime may crash with batch_size > 1. This method always processes exactly one input.

Errors

Returns ClassifierError if tokenization or inference fails.

Note

This method requires &mut self due to ort 2.0 API requirements. Use a Mutex wrapper if concurrent access is needed.

pub fn model_version(&self) -> &str

Get the model version.