Skip to main content

​
Overview

Use your on-prem Blast deployment to score real production conversations with evaluator policies you already have in the platform. Send multiple conversations in one request by adding them to the conversations array. The endpoint queues conversations for scoring and returns their queue status so you can monitor completion in the UI or via API. Endpoint: POST {ON_PREM_BASE_URL}/api/projects/{project_id}/conversations/score Use your own on-prem base URL (e.g., https://blast.yourcompany.com).

​
Request schema

  • conversations (array)
    • conversation_id (string): Your identifier for the conversation.
    • turns (array):
      • user_input (string): What the user said.
      • raw_api_output (object): JSON of the model response you want scored. Include only the fields you want evaluators to inspect (e.g., content, references, products, metadata).
  • evaluator_names (array): Names of evaluators configured in the project (e.g., ["search_term_relevance"]).

​
Minimal example payload

{
  "conversations": [
    {
      "conversation_id": "abc-123",
      "turns": [
        {
          "user_input": "What should I do if paint gets on my skin or clothes?",
          "raw_api_output": {} // JSON of the model response you want evaluated
        }
      ]
    }
  ],
  "evaluator_names": ["search_term_relevance"]
}

​
Example request (cURL)

curl -X POST "$ON_PREM_BASE_URL/api/projects/$BLAST_PROJECT_ID/conversations/score" \
  -H "Authorization: Bearer $JWT_TOKEN" \
  -H "Content-Type: application/json" \
  -d @payload.json
Where payload.json contains the JSON payload above.

​
Response shape

{
  "status": "queued",
  "queued_conversations": 1,
  "failed_conversations": 0,
  "total_events": 1,
  "conversations": [
    {
      "conversation_id": "abc-123",
      "id": "internal-id",
      "status": "queued",
      "events_count": 1,
      "error": null,
      "error_code": null
    }
  ]
}

​
Viewing results

  • UI: {ON_PREM_BASE_URL}/{project_id}/production
  • API: Poll monitoring endpoints if you need programmatic status checks.
Use evaluator names that exist in your project. Ensure the bearer token has access to the target project in your on-prem environment.