Struct HydraModel 

pub struct HydraModel { /* private fields */ }

Hydra model wrapper

Supports multiple inference backends:

1. Native safetensors (preferred) - pure Rust inference
2. Heuristic fallback - rule-based when model unavailable

Tokenizer Support

Hydra uses trait-based tokenization supporting:

• Llama 3 (128K vocab) - default for open source
• OpenAI o200k_base (200K vocab)
• OpenAI cl100k_base (100K vocab)
• Fallback byte-level tokenizer

Vocab Mismatch Handling

The current model (v1.0) was trained with 32K vocab. When using a larger tokenizer (e.g., Llama 3 128K), token IDs exceeding the model’s vocab size are clamped to vocab_size - 1. This preserves functionality but may degrade accuracy. A warning is logged when this occurs.

Implementations

impl HydraModel

pub fn new() -> Self

Create new model (unloaded, with byte-level tokenizer)

pub fn fallback_only() -> Self

Create model with fallback only (no neural inference)

pub fn with_tokenizer(tokenizer: BoxedTokenizer) -> Self

Create model with a specific tokenizer

pub fn load<P: AsRef<Path>>(path: P) -> Result<Self>

Load model from directory or file

Expects directory structure:

./models/hydra/
└── model.safetensors

Or a direct path to .safetensors file.

Falls back to heuristics if loading fails.

pub fn is_loaded(&self) -> bool

Check if model is loaded

pub fn model_path(&self) -> Option<&str>

Get the model path (if loaded or attempted to load)

pub fn uses_fallback(&self) -> bool

Check if fallback heuristics are enabled

pub fn tokenizer_type(&self) -> TokenizerType

Get the tokenizer type

pub fn vocab_size(&self) -> usize

Get vocabulary size

pub fn model_vocab_size(&self) -> usize

Get the model’s vocabulary size (may differ from tokenizer)

pub fn has_vocab_mismatch(&self) -> bool

Check if there’s a tokenizer/model vocab mismatch

pub fn predict_compression(&self, content: &str) -> Result<CompressionDecision>

Predict compression algorithm for content

pub fn predict_security(&self, content: &str) -> Result<SecurityDecision>

Predict security status for content

Trait Implementations

impl Clone for HydraModel

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Default for HydraModel

fn default() -> Self

Returns the “default value” for a type.