RetrievalKit Embedding for Node.js

@gungorbasa/retrievalkit-embedding is RetrievalKit's optional Node.js embedding package. It produces local FP32 MiniLM embeddings through the official ONNX Runtime 1.24.3 without adding embedding dependencies to the base retrieval package.

The production contract is fixed: at most 256 tokens and exactly 384 finite, L2-normalized Float32Array values. RetrievalKit databases continue to accept F32 vectors publicly and may store them with the independent I8ScalarQuantized database encoding.

import { OnnxEmbedder } from "@gungorbasa/retrievalkit-embedding";

await using embedder = await OnnxEmbedder.load({
  runtimeLibraryPath: "/application/lib/libonnxruntime.1.24.3.dylib"
});
const vector = await embedder.embed("local semantic retrieval");

load() and prefetch() are the only APIs that may acquire verified model artifacts. embed() and embedBatch() use an already-loaded session and do not perform network or model-cache access. Set localOnly: true to refuse model downloads.

The application normally supplies the official runtime with runtimeLibraryPath or RETRIEVALKIT_ONNX_RUNTIME_LIBRARY. A repository-local packaging build can opt into copying a verified runtime:

RETRIEVALKIT_BUNDLE_ONNX_RUNTIME=1 \
RETRIEVALKIT_ONNX_RUNTIME_LIBRARY=/path/libonnxruntime.1.24.3.dylib \
npm run build:native

Only the qualified runtime with exact size 27,724,968 bytes and SHA-256 b65e22247d3ce2976931cfc6be3929e6fb81cd55e2f202e95e0ab8c9de5fa729 is accepted for package-local discovery. The binary is not stored in this repository.

The package is included in the v0.1.0 release inventory but is not available from npm until the protected release gates pass.