api-service-sac/granite-embedding-97m-multilingual-r2-onnx

granite-embedding-97m-multilingual-r2 (ONNX)

An ONNX export of ibm-granite/granite-embedding-97m-multilingual-r2 by IBM, packaged for s1grep. s1grep uses it for its quick first pass over a large repository: it embeds the outline of every function (path, name and first lines) several times faster than the larger retriever, so the whole repository can be searched within minutes while whole sources are still being indexed.

• model.onnx: FP32 graph with dynamic batch and sequence axes; its output is last_hidden_state.
• embedder_config.json: how s1grep turns that output into a vector (CLS pooling, L2 normalisation, 512 tokens).
• tokenizer.json: the original tokenizer.

The export matches sentence-transformers to a cosine similarity of 0.999 or more on s1grep's parity tests. The weights are unchanged; all credit for the model goes to IBM.

License

Apache 2.0, the licence of the original model. Copyright IBM.