Module onnx 

ONNX Runtime local embedding provider.

Provides local embedding inference using ONNX Runtime, eliminating the need for an external API. Supports sentence-transformer models such as all-MiniLM-L6-v2 exported to ONNX format.

Feature gating

When compiled without the onnx feature the module provides a stub that validates the model path but returns Error::Embedding from embed() and embed_batch().

When compiled with the onnx feature the module loads the ONNX session and a HuggingFace tokenizer, then performs real local inference with mean-pooling and L2 normalisation.

[features]
onnx = ["dep:ort", "dep:tokenizers", "dep:ndarray"]

[dependencies]
ort = { version = "2.0.0-rc.11", optional = true }
tokenizers = { version = "0.23", optional = true, default-features = false, features = ["fancy-regex"] }
ndarray = { version = "0.17", optional = true }

These are the versions the workspace actually pins. The #[cfg(feature = "onnx")] inference path is written against — and builds + tests against — this ort 2.0.0-rc.11 / ndarray 0.17 / tokenizers 0.23 API (the migration that repaired the old ndarray-0.16 drift). A dedicated onnx feature CI job (.github/workflows/ci.yml) builds and tests --features onnx so it cannot silently rot; it stays out of the workspace-wide jobs only because ort is a heavy native dependency. The one open item on https://github.com/sattyamjjain/mnemo/issues/125 is a model-fetch CI job to make the ONNX MiniLM recall number itself reproducible (end-to-end inference needs a real model on disk via MNEMO_ONNX_MODEL_PATH, which the build+test job does not fetch). Build locally with --features onnx.

Example (stub)

use mnemo_core::embedding::onnx::OnnxEmbedding;
use mnemo_core::embedding::EmbeddingProvider;

// Will succeed only if the path exists on disk.
let provider = OnnxEmbedding::new("/models/all-MiniLM-L6-v2.onnx", 384)
    .expect("model path must exist");

assert_eq!(provider.dimensions(), 384);
assert_eq!(provider.model_path(), "/models/all-MiniLM-L6-v2.onnx");

Structs

OnnxEmbedding

ONNX-based local embedding provider.