Skip to main content

Model Overview

This project provides an MLflow-compatible model that acts as a proxy to an external backend service. Instead of performing local machine learning inference, the model forwards user queries to a backend API and returns structured responses.

Request Format​

The model expects input in JSON format following the MLflow scoring protocol.

Example Request​

{
  "inputs": [
    {
      "question": "Which documents mention institutions or universities?",
      "dataset": "research_production",
      "agent": "domain"
    }
  ]
}

Field Descriptions​

  • inputs (array) – A list of one or more query objects (each representing a single request).
  • question (string) – The user’s natural language query.
  • dataset (string) – Specifies which dataset to use (e.g. research_production, ai_ecosystem).
  • agent (string) – Selects which backend agent handles the request (e.g. domain, data_exploration).

Response Format​

Example Response​

{
  "content": {
    "images": [],
    "text": "Documents mentioning institutions or universities are found in Articles 37, 56, and 68."
  },
  "explanation": {
    "code": "",
    "sources": [
      {
        "article": "Article 37",
        "page": 20,
        "title": "Law on Gender Equality",
        "url": "https://example.com/doc1.pdf"
      }
    ]
  }
}

Field Descriptions​

  • content (object) – The main user-facing answer.

    • text (string) – The generated natural language response.
    • images (array) – Optional list of images related to the answer.

  • explanation (object) – Supporting details for transparency and traceability.

    • code (string) – Optional code used to produce the answer (mainly for data queries).

    • sources (array) – List of references backing the answer:

      • article (string) – Referenced article or section.
      • page (number) – Page number in the document.
      • title (string) – Document title.
      • url (string) – Link to the source document.

Architecture​

Client → MLflow → ModelMapper → Model → External API

Usage​

mlflow models serve -m "models:/social_policy_making_model/1" --env-manager local -p 5001

Registering running model on Databricks​

Add the deployment​

curl --location 'https://staging.private.tango.u-hopper.com/api/deployment' \
--header 'Content-Type: application/json' \
--header 'Authorization: Bearer YOUR_TOKEN' \
--header 'Cookie: language=en-us' \
--data '{
    "deployment_key":"afl-dbx_social-policy-making-model",
    "registry_connector_class":"tango_shared.connectors.deployment.databricks_registry_connector.DatabricksModelRegistryConnector",
    "registry_connector_init":{
        "model_registry_url": "https://YOUR_DBX_INSTANCE.azuredatabricks.net/",
        "tracking_server_url": "https://YOUR_DBX_INSTANCE.azuredatabricks.net/",
        "tracking_server_pat": "YOUR_DBX_PAT"
    },
    "server_proxy_class":"tango_shared.connectors.deployment.databricks_server_proxy.DatabricksModelServerProxy",
    "server_proxy_init":{
        "server_proxy_url": "https://YOUR_DBX_INSTANCE.azuredatabricks.net/serving-endpoints/social_policy_making_model",
        "server_proxy_pat": "YOUR_DBX_PAT"
    }
}'

Add the model​

curl --location 'https://staging.private.tango.u-hopper.com/api/model' \
--header 'Content-Type: application/json' \
--header 'Authorization: Bearer YOUR_TOKEN' \
--header 'Cookie: language=en-us' \
--data '{
    "deployment_key": "afl-dbx_social-policy-making-model",
    "model_code": "social_policy_making_model_mapper",
    "name": "Social Policy Making model on DBX Workspace Model Registry",
    "description": "This model providing features for T5.5 Social Policy Making case study",
    "version": "2",
    "realtime_execution": false,
    "allow_download": false,
    "need_context": false,
    "model_card_file_id": "YOUR_MODEL_CARD_FILEID",
    "input_schema": {
        "fields": [{"type": "string", "name": "question", "required": "True"},
              {"type": "string", "name": "dataset", "required": "True"},
              {"type": "string", "name": "agent", "required": "True"}]
    },
    "output_schema": {
        "fields": [{"type": "string", "name": "content", "required": "True"}]
    }
}'

Approve the model​

curl --location --request PATCH 'https://staging.private.tango.u-hopper.com/api/model/social_policy_making_model_mapper/version/1' \
--header 'Content-Type: application/json' \
--header 'Authorization: Bearer YOUR_TOKEN' \
--header 'Cookie: language=en-us' \
--data '{
    "status": "APPROVED"
}'