AI Platform Data and Contracts

MongoDB prediction collection

The controller stores one document per prediction in collection prediction.

Mongo stores these statuses as integers:

• answer: 0 NO_ANSWER, 1 TRANSCRIBING, 2 TRANSCRIBE_SUCCESS, 3 CONVERT_FILE_FAILED
• result item: 0 WAITING_FOR_ANSWER, 1 PREDICTING, 2 PREDICT_SUCCESS, 3 PREDICT_FAILED

The Go GPT worker can store the submitted answer map, provider response, endpoint, version, and task metadata in openai-gpt-response. Release v1.0.3 enables this option by default, and the production-v2 task definition has no disabling override.

S3 object layout

The controller uploads source media beneath:

<prediction_id>/<question_id>/<original_filename>

The converter output path uses the same prefix with converted_<text-before-first-period>.mp3. Production-v2 uses bucket dmind-aiservice-asrfilestorage-prod-v2.

FIFO queue contracts

ASR request

Queue: dmind_ai-service_asr_queue_prod-v2.fifo.

{
  "filepath": "<sanitized-s3-key>",
  "prediction_id": "<uuid>",
  "question_id": "qs-01-02"
}

The on-cloud, on-prem caller, and OpenAI ASR workers are competing consumers of this queue. The controller does not route a message to a specific implementation; every worker publishes transcripts to the controller result queue. Tier-1 and tier-2 ASR dead-letter queues handle failed processing.

DSC request

Queue: dmind_ai-service_dsc_queue_prod-v2.fifo.

{
  "prediction_id": "<uuid>",
  "text": {
    "qs-01-02": "<sanitized-text>"
  }
}

The enabled event-source mapping targets alias distributer of dmind_ai-service_dsc_invoker-function_prod-v2; that alias points to function version 7. The Lambda calls the DSC model API and emits depression.rank to the controller result queue.

OpenAI request

Queue: dmind_ai-service_openai_queue_prod-v2.fifo.

{
  "prediction_id": "<uuid>",
  "version": "1_0_0",
  "answer": {
    "qs-07-03": "<sanitized-text>"
  },
  "task_id": "openai_suicidal_severity"
}

The Go worker formats the versioned prompt, calls Azure OpenAI, parses function-call or content JSON into task-specific structures, and sends normalized result fragments.

Controller result

Queue: dmind_ai-controller_result-queue_prod-v2.fifo.

The SQS message attribute type selects one of two body contracts:

• store_text: prediction_id, question_id, and transcript text.
• save_result: prediction_id and an array of {result_id, value}.

Successful persistence removes the message. Malformed JSON and unrecognized non-empty type values are also deleted; persistence errors and messages without type remain for another receive.

Result contract

The internal Mongo result tree is versioned and can aggregate:

• two-question-plus flags: sadness, boredom, suicidal thought
• depression: numeric/model rank, choice rank where applicable, and LOW/MODERATE/SEVERE level
• suicidal assessment: severity, rank, and detail
• chief complaint: sad complaint terms

The public GET /result envelope contains prediction id, status (PROCESSING or FINISH), optional result, and answer_count. A finished result exposes:

Field	Public type
two_q_plus	{sad: boolean, boredom: boolean, suicidal_thought: boolean}
suicidal_score	integer in 1_0_0; integer or null in 1_1_0 and 1_2_0
chief_complaint	{sad: string[]}
depression_level	LOW, MODERATE, or SEVERE

Versions 1_1_0 and 1_2_0 finish against method-specific required results; 1_0_0 requires the complete legacy result set. Version 1_2_0 derives two_q_plus.suicidal_thought from suicidal rank. No public version exposes the full internal depression or suicidal tree. Clients depend on the requested result-format version, not the Mongo representation.