Features in Inference

This tutorial shows how to access data stored in the JFrThis tutorial shows how to access data stored in the JFrog ML Feature Store online store during the inference.

Using `OnlineClient`

In the predict function, we create an instance of the OnlineClient.

After that, we create a ModelSchema containing all of the features we want to retrieve.

We create a DataFrame containing the entities identifiers and pass it to the Feature Store. As a response, we get a Pandas DataFrame with the requested features.

import pandas as pd
from frogml.feature_store.online.client import OnlineClient
from frogml.sdk.model.schema_entities import FeatureStoreInput
from frogml.sdk.model.schema import ModelSchema

model_schema = ModelSchema(
    inputs=[
        FeatureStoreInput(name='user-credit-risk-features.checking_account'),
        FeatureStoreInput(name='user-credit-risk-features.age'),
        FeatureStoreInput(name='user-credit-risk-features.job'),
        FeatureStoreInput(name='user-credit-risk-features.duration'),
        FeatureStoreInput(name='user-credit-risk-features.credit_amount'),
        FeatureStoreInput(name='user-credit-risk-features.housing'),
        FeatureStoreInput(name='user-credit-risk-features.purpose'),
        FeatureStoreInput(name='user-credit-risk-features.saving_account'),
        FeatureStoreInput(name='user-credit-risk-features.sex'),
        FeatureStoreInput(name='liked-posts.count')
    ])

online_client = OnlineClient()

df = pd.DataFrame(columns=['user', 'post_id'],
                  data=[['06cc255a-aa07-4ec9-ac69-b896ccf05322', '1234'],
                        ['asdc255a-aa07-4ec9-ac69-b896c1231445', '7889']])

user_features = online_client.get_feature_values(model_schema, df)

print(user_features)

Using the `frogml.api()` Decorator

Alternatively, we could use the features_extraction parameter and get the features automatically extracted when the Entity is being sent in your prediction input Dataframe . As with the OnlineClient, the ModelSchema is required to define what features are to be extracted from the Online Store.

# model.py
import frogml

@frogml.api(feature_extraction=True)
def predict(self, df, extracted_df):
  
    # Add prediction logic here
    
    return output_dataframe

In the predict function, use the frogml.api()decorator with the parameter feature_extraction=True as in the code example above.

The df Dataframe will contain your inference call inputs and the extracted_df Dataframe will contain the latest feature vectors from the Online store for the queried entities.

Using the REST API

The JFrog ML Online Store can also be queried via REST calls as shown in the example below:

Generate a JFrog ML Token for Authentication

Use the following command to obtain a token, valid for 24 hours:

curl --request POST 'https://grpc.qwak.ai/api/v1/authentication/qwak-api-key' \
      --header 'Content-Type: application/json' \
      --data '{"qwakApiKey": "<QWAK_API_KEY>"}'

Optionally, store the token in an environment variable:

export JFROG_TOKEN="<OUTPUT_FROM_AUTHENTICATION_CURL_CALL>"

Retrieve Online Features

With the fresh JFrog ML token, use the following command to extract features:

curl --location 'https://grpc.<YOUR-ACCOUNT>.qwak.ai/api/v1/rest-serving/multiFeatureValues/' \
--header 'Authorization: Bearer '$JFROG_TOKEN'' \
--header 'Content-Type: application/json' \
--data '{
  "entitiesToFeatures": [{
    "features": [{
      "batchV1Feature": {
        "name": "user-credit-risk-features.checking_account"
      }
    }, {
      "batchV1Feature": {
        "name": "user-credit-risk-features.age"
      }
    }],
    "entityName": "user"
  }],
  "entityValuesMatrix": {
    "header": {
      "entityNames": ["user"]
    },
    "rows": [{
      "entityValues": ["45b7836f-bf7c-4039-bc9e-d33982cc1fc5"]
    }]
  }
}'

Note

When referencing feature sets in SDK or REST calls, use hyphens - instead of underscores _. This is a common notation in the JFrog ML platform to ensure consistency and avoid errors during calls.

Important

If you're on a SaaS account, use batchV1Feature as suggested in the example above.

For hybrid accounts, switch to using batchFeature in the REST data JSON payload.

Example JSON Response:

{
  "featureValues": "{\"index\":[0],\"data\":[[\"moderate\",27]],\"columns\":[\"user-credit-risk-features.checking_account\",\"user-credit-risk-features.age\"]}"
}

These examples are using curl for REST calls but any other REST client will work just as well.

Using OnlineClient