Support Create If Not Exists for Tensor Cells in Tensor Field Modifications

[karowan]

Is your feature request related to a problem? Please describe. A clear and concise description of what the problem is. Ex. I'm always frustrated when [...]

It would be great if there were a way to create a tensor cell if it does not exist with a tensor modification request during a tensor partial update via the document api. Currently if you do a modify operation on a tensor field and the cell you wish to modify does not exist, the modification request is ignored for that cell and thus, even when using the create=true parameter on the document request, the document is created but with an empty tensor field.

Describe the solution you'd like A clear and concise description of what you want to happen.

I propose a parameter in the document api PUT which would allow a cell to be automatically created with reasonable defaults such that a modification request would result in the originally non-existent field to be populated with the expected values.

Additional context Add any other context or screenshots about the feature request here.

As a more concrete example of what I would like, assume we have the following document schema within the content cluster named "content":

schema tensor_document {

    document tensor_document {
        field tensor_field type tensor<float>(x{}) {
            indexing: attribute
        }
    }
}

Currently if I were to make the following request:

PUT /document/v1/content/tesnor_document/docid/1234?create=true HTTP/1.1
Content-Type: application/json

{
    "fields": {
        "tensor_field": {
            "modify": {
                "operation": "add",
                "cells": {
                    "a": 1.0
                }
            }
        }
    }
}

I would get back the following from a GET request to the same url

{
    "pathId": "/document/v1/content/tensor_document/docid/1234",
    "id": "id:content:tensor_document::1234",
    "fields": {}
}

What I would like is for the fields to contain a tensor with the cell a:1.0 similarly to how it would exist if you were to use add instead of modify in the document update request.

[vekterli]

The semantics of create=true only apply at the document level and not the field (or cell) level. I think for a feature like this we'd likely want a separate modify operation with cell upsert semantics (or enhance the semantics of an existing operation), as these explicitly operate on specific fields. This could also be used together with create=true to effectively create a single-cell document if none already existed.

[bratseth]

Yes. We could e.g signal this by passing a default value with the operation. The alternative is to use 1 with multiplicative and 0 with additive ops.

[geirst]

Support for "create": true on tensor modify update is available in Vespa 8.217.43 (https://github.com/vespa-engine/vespa/pull/28152). System test: https://github.com/vespa-engine/system-test/pull/2988 Reference documentation: https://github.com/vespa-engine/documentation/pull/2838

[geirst]

In Vespa 8.218.31, "create": true also works for empty (non-existing) tensor fields (https://github.com/vespa-engine/vespa/pull/28186). System test: https://github.com/vespa-engine/system-test/pull/2989

[geirst]

Vespa 8.218.31 is now available. Reference doc: https://docs.vespa.ai/en/reference/document-json-format.html#tensor-modify