Struct ModelManager 

pub struct ModelManager { /* private fields */ }

Manages downloading, caching, and verifying ONNX model files.

Implementations

impl ModelManager

pub fn new() -> Result<Self, ModelManagerError>

Create a new ModelManager using the default cache directory (~/.cache/ctxgraph/models/).

pub fn with_cache_dir(cache_dir: PathBuf) -> Result<Self, ModelManagerError>

Create a ModelManager with a custom cache directory (useful for tests).

pub fn default_cache_dir() -> Result<PathBuf, ModelManagerError>

Return the default cache directory (~/.cache/ctxgraph/models/), creating it if it does not exist.

pub fn model_path(&self, spec: &ModelSpec) -> PathBuf

Path where a given model would be stored locally.

pub fn is_cached(&self, spec: &ModelSpec) -> bool

Check whether the model file exists on disk and its size matches the spec.

pub fn verify(&self, spec: &ModelSpec) -> Result<bool, ModelManagerError>

Verify the SHA-256 hash of a cached model file. Returns Ok(true) if the hash matches, Ok(false) if it doesn’t, or an error if the file cannot be read.

If spec.sha256 starts with “pending” or equals “skip”, verification is bypassed and Ok(true) is returned unconditionally.

pub fn download(&self, spec: &ModelSpec) -> Result<PathBuf, ModelManagerError>

Download a model, verify its hash, and return the local path.

pub fn get_or_download( &self, spec: &ModelSpec, ) -> Result<PathBuf, ModelManagerError>

Return the cached model path if it exists and is valid, otherwise download it.

impl ModelManager

pub fn ensure_ner_models(&self) -> Result<(PathBuf, PathBuf), ModelManagerError>

Download the NER model and tokenizer needed for Tier 1 extraction.

Downloads gliner_large-v2.1 INT8 model + tokenizer to the cache directory.

pub fn ensure_rel_models(&self) -> Option<(PathBuf, PathBuf)>

Download the multitask model and tokenizer needed for relation extraction.

Returns None if the model is not available (needs ONNX conversion).

pub fn ensure_nli_models(&self) -> Result<(PathBuf, PathBuf), ModelManagerError>

Download the NLI cross-encoder model and tokenizer.

Used for zero-shot relation classification via entailment scoring.

pub fn find_nli_model(&self) -> Option<(PathBuf, PathBuf)>

Find locally cached NLI model. Returns None if not downloaded yet.

pub fn find_relation_classifier(&self) -> Option<PathBuf>

Check for a locally available fine-tuned relation classifier model.

Looks for relation_classifier/model_int8.onnx (or model.onnx) and relation_classifier/tokenizer.json in the cache directory.

Returns Some((model_path, tokenizer_path)) if found, None otherwise.

pub fn find_relex_model(&self) -> Option<(PathBuf, PathBuf)>

Check for locally exported gliner-relex ONNX model.

The relex model must be exported manually using: python scripts/export_relex_onnx.py [--quantize]

Returns Some((model_path, tokenizer_path)) if found, None otherwise.