Method: projects.locations.collections.dataStores.trainCustomModel

Trains a custom model.

HTTP request

POST https://discoveryengine.googleapis.com/v1/{dataStore=projects/*/locations/*/collections/*/dataStores/*}:trainCustomModel

The URL uses gRPC Transcoding syntax.

Path parameters

Parameters
dataStore

string

Required. The resource name of the Data Store, such as projects/*/locations/global/collections/default_collection/dataStores/default_data_store. This field is used to identify the data store where to train the models.

Request body

The request body contains data with the following structure:

JSON representation 
{
  "modelType": string,
  "errorConfig": {
    object (ImportErrorConfig)
  },
  "modelId": string,

  // The following is a list of mutually exclusive fields. At most one of the
  // fields will be set in a response:
  "gcsTrainingInput": {
    object (GcsTrainingInput)
  }
  // End of mutually exclusive fields.
}
Fields
modelType

string

Model to be trained. Supported values are:

  • search-tuning: Fine tuning the search system based on data provided.
errorConfig

object (ImportErrorConfig)

The desired location of errors incurred during the data ingestion and training.
modelId

string

If not provided, a UUID will be generated.
Model training input. The following is a list of mutually exclusive fields. At most one of the fields will be set in a response:
gcsTrainingInput

object (GcsTrainingInput)

Cloud Storage training input.
End of mutually exclusive fields.

Response body

If successful, the response body contains an instance of Operation.

Authorization scopes

Requires one of the following OAuth scopes:

  • https://www.googleapis.com/auth/cloud-platform
  • https://www.googleapis.com/auth/discoveryengine.readwrite
  • https://www.googleapis.com/auth/discoveryengine.serving.readwrite

For more information, see the Authentication Overview.

IAM Permissions

Requires the following IAM permission on the dataStore resource:

  • discoveryengine.dataStores.trainCustomModel

For more information, see the IAM documentation.

GcsTrainingInput

Cloud Storage training data input.

JSON representation 
{
  "corpusDataPath": string,
  "queryDataPath": string,
  "trainDataPath": string,
  "testDataPath": string
}
Fields
corpusDataPath

string

The Cloud Storage corpus data which could be associated in train data. The data path format is gs://<bucket_to_data>/<jsonl_file_name>. A newline delimited jsonl/ndjson file.

For search-tuning model, each line should have the Id, title and text. Example: {"Id": "doc1", title: "relevant doc", "text": "relevant text"}
queryDataPath

string

The gcs query data which could be associated in train data. The data path format is gs://<bucket_to_data>/<jsonl_file_name>. A newline delimited jsonl/ndjson file.

For search-tuning model, each line should have the Id and text. Example: {"Id": "query1", "text": "example query"}
trainDataPath

string

Cloud Storage training data path whose format should be gs://<bucket_to_data>/<tsv_file_name>. The file should be in tsv format. Each line should have the docId and queryId and score (number).

For search-tuning model, it should have the query-id corpus-id score as tsv file header. The score should be a number in [0, inf+). The larger the number is, the more relevant the pair is. Example:

  • query-id\tcorpus-id\tscore
  • query1\tdoc1\t1
testDataPath

string

Cloud Storage test data. Same format as trainDataPath. If not provided, a random 80/20 train/test split will be performed on trainDataPath.