Custom Taxonomy Classification API

API on Google Cloud Run to classify text, images and videos into a custom taxonomy.

The API exposes endpoints to:

• Use Google's embedding models to generate embedding vectors for a provided list of categories, create an index using tree-AH algorithm and expose the index via a Google Cloud Vector Search endpoint (initial setup)
• Classify any given text, image or video using the above index and endpoint.
• Track the status of the initial setup task.

Features:

• Import a taxonomy from a Google Spreadsheet.
• Integration with Google Vector Search for fast nearest neighbor searches.
• Rate limiting for Vertex API calls.
• VPC network with dedicated firewall rules.

Prerequisites

1. The user deploying the application to Google Cloud must be a Project Owner.
2. Have a Google Spreadsheet containing the taxonomy (list of category names).

Installation

To deploy, run the following commands in a Google Cloud Shell and follow the instructions. If there's a problem with the deployment then address the error message and run the command again.

git clone [email protected]:google-marketing-solutions/custom-taxonomy-classifier.git && \
cd custom-taxonomy-classifier/api && \
chmod 775 ./setup.sh && \
./setup.sh

Grant the application service account Viewer access to the Google Spreadsheet containing the taxonomy. Replace the <your-cloud-project-id> with the name of the Google Cloud project you intend to use.

classify-api-sa@<your-cloud-project-id>.iam.gserviceaccount.com

API Usage

The API has 3 endpoints:

1. POST /generate_taxonomy_embeddings

   Used to generate and store the vector embeddings for each of the taxonomy nodes. This will run as a background process and returns a task_id.

   NOTE: This endpoint needs to be called at least once before the other endpoints below are functional. Calling this endpoint overwrites any existing Vector Search endpoints.

2. GET /task_status/{task_id}

   Used to get the current status of the background process to generate and store the taxonomy embeddings vectors.

3. POST /classify

   Used to get the top 10 categories in terms of embeddings vector similarity for the passed text. The below graph illustrates the architecture for the classify call:

Generate Taxonomy Embeddings

The endpoint triggers a background process to get the category names from a Google Spreadsheet, generates their embeddings and writes them to a postgres database. Calling the endpoint creates a task, which can be used to retrieve the task status using the task_status endpoint.

Request format

Parameters

• spreadsheet_id (required): The ID of the Google Spreadsheet.
• worksheet_name (required): The name of the worksheet that contains the taxonomy.
• worksheet_col_index (required): The 1-based column index of the column containing the list of categories.
• header (optional): Whether or not the column has a header row. Defaults to true.

curl -X POST \
-H "Authorization: Bearer $(gcloud auth print-identity-token)" \
-H 'Content-Type: application/json' \
-d '{"spreadsheet_id": "[YOUR-GOOGLE-SPREADSHEET-ID]", "worksheet_name": "Sheet1", "worksheet_col_index": "1", "header": "False"}' \
-i [YOUR-CLOUD-RUN-URL]/generate_taxonomy_embeddings

Response format

The API response will return a task_id including a message.

Attributes

• task_id: The ID of the task created by the background process.
• message: An informational message about the task creation.

{
  "task_id": "9687244b-6883-474a-97f6-a29f5f91b522",
  "message": "Generate Taxonomy Embeddings task sent in the background."
}

Task Status

The endpoint can be used to retrieve the current status of a particular task.

Request format

Parameters

• task_id (required): The ID of the Google Spreadsheet.

curl -X GET \
-H "Authorization: Bearer $(gcloud auth print-identity-token)" \
-i [YOUR-CLOUD-RUN-URL]/task_status/9687244b-6883-474a-97f6-a29f5f91b522

Response format

The API response will return a task_id including a message.

Attributes

• task_id: The ID of the task created by the background process.
• status: The status of the task.
• created_time: The time the task was created.
• updated_time: The time the task was last updated.
• message: An informational message about the task. Is empty by fault but will be populated on error.

{
  "task_id": "9687244b-6883-474a-97f6-a29f5f91b522",
  "status": "SUCCESS",
  "created_time": "2024-01-17T15:07:52.633122Z",
  "created_updated": "2024-01-17T15:11:42.954784Z",
  "message": null
}

Classify

The endpoint returns a list of the top 10 best matching categories, along with their corresponding scores (similarity).

Request format

Parameters

• text (optional): The text content to classify. This can be a single string or a list of strings.
• media_uri (optional): The media to classify. This can be a single file path to a GCS location or a list of file paths. Note that only the following media extensions are supported: .jpg, .jpeg, .png, .x-flv, .mov, .mpeg, .mpegps, .mpg, .mp4, .webm, .wmv, .3gpp
• embeddings (optional): Whether or not to include the generated embeddings for text and media. The default values is false.

With a list of strings:

curl -X POST \
-H "Authorization: Bearer $(gcloud auth print-identity-token)" \
-H 'Content-Type: application/json' \
-d '{"text": ["Text to classify","Some other Text to classify"]}' \
-i [YOUR-CLOUD-RUN-URL]/classify

With a single string:

curl -X POST \
-H "Authorization: Bearer $(gcloud auth print-identity-token)" \
-H 'Content-Type: application/json' \
-d '{"text": "Text to classify", "embeddings": "true"}' \
-i [YOUR-CLOUD-RUN-URL]/classify

Response format

The API response will show the top 10 categories in terms of dot product similarity to the passed text.

Categories Similarity Object

Attributes:

• name: The name of the category within the given taxonomy.
• similarity: A float of the dot product similarity of the category to the passed text.

Results Object

Attributes:

• text: The text content string that was classified.
• media_uri: The image or video content that was classified.
• categories: An array containing category similarity objects.
• media_description: A description of the passed media object.
• embedding: The embedding vector for the text or media. Only present when the embeddings argument was set to true in the request.

Example Response:

The below would be a response for the following request data:

{
  "text": "Text to classify",
  "media_uri": "gs://path/to/image.jpg"
  "embeddings": "true"
}

[
  {
    "text": "Text to classify"
    "media_uri": "null",
    "media_description": "null",
    "categories": [
      {
        "name": "Some category",
        "similarity": 0.9
      },
      {
        "name": "Some other category",
        "similarity": 0.8
      },
      ...
    ],
    "embedding": [0.1, 0.2, 0.3, ...]
  },
  {
    "text": "null",
    "media_uri": "gs://path/to/image.jpg",
    "media_description": "A cat with blue eyes looking up.",
    "categories": [
      {
        "name": "Some category",
        "similarity": 0.9
      },
      {
        "name": "Some other category",
        "similarity": 0.8
      },
      ...
    ],
    "embedding": [0.3, 0.2, 0.1, ...]
  }
]