@ryanstark24/sfgraph-models

Vendored embedding model for sfgraph. Ships an ONNX-quantised all-MiniLM-L6-v2 (384-dim) under data/ via Git LFS — see the root .gitattributes. Used by @ryanstark24/sfgraph-core to generate node embeddings for semantic search.

Install

The .onnx and tokenizer files are tracked with Git LFS. After cloning:

git lfs install
git lfs pull

Without this the files on disk are 134-byte LFS pointers and loadEmbedder() will fail its SHA-256 checksum verification.

@xenova/transformers is an optionalDependency — installing this package without it disables embedding generation but does not error.

Usage

import { loadEmbedder } from "@ryanstark24/sfgraph-models";

const embedder = await loadEmbedder();
const [vec] = await embedder.embed(["account.status field accessor"]);
// vec is a Float32Array of length 384
await embedder.close();

Bring your own model

Override the vendored model via either options or env vars:

| Option | Env var | Default | | ----------- | -------------------------- | ------------------------- | | modelPath | SFGRAPH_EMBED_MODEL_PATH | <package>/data | | modelId | SFGRAPH_EMBED_MODEL_ID | Xenova/all-MiniLM-L6-v2 | | dim | SFGRAPH_EMBED_MODEL_DIM | 384 | | quantized | — | true |

The model directory layout must match what transformers.js expects: <modelPath>/<modelId>/onnx/model.onnx (and the matching tokenizer files).

Data files

data/MODEL_INFO.md documents the exact origin and SHA-256 of every file checked in under data/. Updating the vendored model requires regenerating src/checksum.ts — see that file for the verification flow.