codira-embedding-onnx

First-party native ONNX Runtime embedding engine plugin for Codira.

The package publishes the onnx engine through the codira.embedding_engines entry-point group. Model artifacts are not bundled; operators provision an ONNX model file and tokenizer file explicitly.

Long inputs are truncated before ONNX Runtime inference. The default limit is max_tokens = 512, matching the common fixed sequence length used by BERT-style embedding exports. Dynamic-shape or longer-context ONNX exports can override the limit through repository, user, or system config:

[plugins.embedding-onnx]
max_tokens = 512

Set max_tokens = 0 only for ONNX exports that are known to accept arbitrary sequence lengths. Changing max_tokens changes vectors for over-limit texts, so bump [embeddings].version when changing it for an existing index.