Adding support for specifying embedding type to Jina AI service settings

[ymao1]

Summary

Adding the ability to specify embedding_type in the Jina AI service settings when creating a Jina AI inference endpoint.

Usage

• Not specifying an embedding_type will default to the float embedding type (same as previous behavior)
• Allowed types: ['float', 'bit', 'binary']
• Specifying embedding_type: 'bit' or embedding_type: 'binary' will return results in the text_embedding_bits key (same as Cohere)

# Create a Jina AI binary inference endpoint
PUT /_inference/text_embedding/jina_embeddings_bit
{
    "service": "jinaai",
    "service_settings": {
        "api_key": <apiKey>,
        "model_id": "jina-embeddings-v3",
        "embedding_type": "bit"
    }
}

# Perform an inference task
POST /_inference/text_embedding/jina_embeddings_bit
{
    "input": "hello",
    "task_settings": {
        "input_type": "ingest"
    }
}

# Response
{
    "text_embedding_bits": [
        {
            "embedding": [
                -55,
                74,
                101,
                67,
                83,
                1,
                53,
                -101,
                -71,
                -98,
                -116,
                -99,
                80,
                -49,
                65,
                .
                .
                .
            ]
        }
    ]
}

Notes

Same as Cohere, requesting a binary embedding from Jina AI will returns an array of binary embeddings packed as bytes with int8 precision. Since this aligns with what is expected as input when you specify a dense_vector mapping with element_type: bit, we do not perform any bit unpacking on the response and handle the bytes as-is.

[elasticsearchmachine]

Hi @ymao1, I've created a changelog YAML for you.

[elasticsearchmachine]

Pinging @elastic/ml-core (Team:ML)

[jonathan-buttner]

Since we use this if logic a few times how about we move it into its own function?

Something like isEmbeddingTypeSupported(TransportVersion version) or something like that.

[ymao1]

💚 All backports created successfully

Status	Branch	Result
✅	8.x

Questions ?

Please refer to the Backport tool documentation