Metadata-Version: 2.4
Name: codira-embedding-onnx
Version: 1.55.0
Summary: First-party native ONNX Runtime embedding engine plugin for codira
Requires-Python: >=3.13
Description-Content-Type: text/markdown
Requires-Dist: codira<2.0.0,>=1.5.0
Requires-Dist: numpy<3.0,>=2.0
Requires-Dist: onnxruntime<2.0,>=1.18
Requires-Dist: tokenizers<1.0,>=0.19
Provides-Extra: test
Requires-Dist: black; extra == "test"
Requires-Dist: mypy; extra == "test"
Requires-Dist: pytest; extra == "test"
Requires-Dist: ruff; extra == "test"

# 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:

```toml
[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.
