Document how custom feature stores are used at query time

[fiorinin]

Hi,

It looks like reranking won't work if we use non-default stores. That is, after I uploaded my model, I can see it at this address:
http://localhost:9203/_ltr/store_name/_model/model_name

Similarly, because it's in a specific store, this address doesn't return anything (per the documentation):
http://localhost:9203/_ltr/_model/

But this one does:
http://localhost:9203/_ltr/store_name/_model/

Then when I need to rescore using

{
    "query": {
        "sltr": {
                "params": {
                    "keywords": query_string
                },
                "model": "model_name",
            }
    }
}

The plugin returns it can't find the model. I tried concatenating the store in the name, or using the ID, vainly. Is there a way, currently, to make this work? If not, could you point me where I should look in the plugin to fix it?

Thanks a lot!

[nomoa]

Sure, you simply have to provide the store name in the store param:

{
    "query": {
        "sltr": {
                "params": {
                    "keywords": query_string
                },
                "model": "model_name",
                "store": "store_name"
            }
    }
}

I haven't tested so feel free to reopen if this does not work as expected.

[ebernhardson]

Sounds like, at minimum, we might need a documentation patch? Will need to review the docs to double check.

[fiorinin]

Thanks a lot! I’ll try that and let you know if there’s a problem. Yes, during my discovery of the plugin - for which I’m very thankful BTW - I previously noticed aspects that weren’t covered in the documentation. I could find the answers by browsing the code usually, but I got stuck on this one. I’m pretty sure I’m missing so many things because the documentation doesn’t seem to reflect how powerful the plugin is.

Thanks a lot again for your answer!

[nomoa]

Thanks, and sorry to have closed this issue too quickly, I agree with @ebernhardson a doc update to mention how to use feature stores is indeed the least we should do.