Code to reproduce https://github.com/vespa-engine/pyvespa/issues/327

pyvespa_onnx_code.py

# clone pyvespa repo: git clone git@github.com:vespa-engine/pyvespa.git
# Install pyvespa from master branch - pip install -e .[full]
# save model.onnx file on the working directory

from vespa.package import ApplicationPackage, Field, OnnxModel, QueryTypeField, RankProfile, Function, FieldSet, SecondPhaseRanking

#
# Create the application package - it assumes you have model.onnx file on the working directory
#
app_package = ApplicationPackage(name="crossencoder")
app_package.query_profile_type.add_fields(
    QueryTypeField(
        name="ranking.features.query(text_embedding)",
        type="tensor<foat>(x[1], y[512])"
    )
)
app_package.schema.add_fields(
    Field(name="temp", type="string", indexing=["index"]),
    Field(
        name="video_embedding", 
        type="tensor<float>(x[1], y[12], z[512])", 
        indexing=["attribute"]
    ),
)
app_package.schema.add_field_set(FieldSet(name="default", fields=["temp"]))
app_package.schema.add_model(
    OnnxModel(
        model_name="cross_encoder", 
        model_file_path="model.onnx", 
        inputs={
            "video_embedding": "attribute(video_embedding)", 
            "text_embedding": "query(text_embedding)"
        }, 
        outputs={"similarity": "similarity"}
    )
)

app_package.schema.add_rank_profile(
    RankProfile(
        name="default", 
        first_phase="similarity",
        second_phase=SecondPhaseRanking(expression="similarity", rerank_count=10),
        functions=[Function(name="similarity", expression="onnx(cross_encoder).similarity{d0:0,d1:0}")],
        summary_features=[
            "similarity", 
            "attribute(video_embedding)", 
            "query(text_embedding)"
        ]
    )
)

#
# Deploy the application package
#
from vespa.deployment import VespaDocker

vespa_docker = VespaDocker()
app = vespa_docker.deploy(application_package=app_package)