Moderation

Classify text for safety with your organization’s self-hosted BRAHMA model. POST /api/v1/voice-engine/llm/v1/moderations The request and response bodies follow the OpenAI moderations schema, so existing OpenAI clients work by pointing at this base URL and authenticating with your PeachDesk API key (X-API-Key). brahma-moderation is a thin classification layer over your organization’s self-hosted BRAHMA LLM plane — the same GPU that serves chat completions — never an external provider. Every response carries per-organization rate-limit headers; non-2xx responses use the OpenAI error format.

Model

Model	Purpose
`brahma-moderation`	Classify text against the 13 policy categories below.

Any model you send is accepted and the response always reports brahma-moderation.

Backends

brahma-moderation can be served by either of two self-hosted backends. The request and response shape is identical for both — the choice is an operator deployment setting, swapped silently without any client change:

Backend	How it classifies
BRAHMA classifier (default)	Your organization’s self-hosted BRAHMA LLM scores the text against all 13 categories in one pass.
BRAHMA-Guardian (optional)	A dedicated, self-hosted safety model tuned for risk detection — including prompt-injection and jailbreak attempts — mapped onto the same 13 categories.

Both run entirely on your own GPUs; nothing is sent to a third-party moderation service. Which backend is active is transparent to your code — you always send and receive the schema below.

Category	Flags content that…
`harassment`	Harasses, bullies, or demeans a person or group.
`harassment/threatening`	Harassment that includes threats of violence or serious harm.
`hate`	Expresses hate toward a protected group.
`hate/threatening`	Hateful content that also threatens violence or serious harm.
`illicit`	Gives advice or instructions for wrongdoing.
`illicit/violent`	Illicit content that references violence or procuring weapons.
`self-harm`	Depicts or encourages self-harm (suicide, cutting, eating disorders).
`self-harm/intent`	Expresses intent to engage in self-harm.
`self-harm/instructions`	Gives instructions or advice for committing self-harm.
`sexual`	Sexual content or content meant to arouse.
`sexual/minors`	Sexual content that involves a minor.
`violence`	Depicts death, violence, or physical injury.
`violence/graphic`	Depicts death, violence, or injury in graphic detail.

Request fields

Field	Type	Notes
`input` (required)	string · string[] (≤ 32)	The text(s) to classify. Batch inputs map to `results[]` in order. Max 50,000 characters per input.
`model`	string	Optional; ignored — always classified by `brahma-moderation`.

Response

{
  "id": "modr-2f6c9b1e8a4d47c0b3e5a1f7c9d02b48",
  "model": "brahma-moderation",
  "results": [
    {
      "flagged": true,
      "categories": {
        "harassment": false,
        "harassment/threatening": false,
        "hate": true,
        "hate/threatening": false,
        "illicit": false,
        "illicit/violent": false,
        "self-harm": false,
        "self-harm/intent": false,
        "self-harm/instructions": false,
        "sexual": false,
        "sexual/minors": false,
        "violence": false,
        "violence/graphic": false
      },
      "category_scores": {
        "harassment": 0.02,
        "harassment/threatening": 0.0,
        "hate": 0.91,
        "hate/threatening": 0.08,
        "illicit": 0.0,
        "illicit/violent": 0.0,
        "self-harm": 0.0,
        "self-harm/intent": 0.0,
        "self-harm/instructions": 0.0,
        "sexual": 0.01,
        "sexual/minors": 0.0,
        "violence": 0.03,
        "violence/graphic": 0.0
      }
    }
  ]
}

results[i] corresponds to input[i]. flagged is true when any category is true or any category_scores value exceeds the server threshold (default 0.5).

Example

curl https://api.stage.peachdeskai.com/api/v1/voice-engine/llm/v1/moderations \
  -H "X-API-Key: $PEACHDESK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": ["I love this product", "you are worthless"]
  }'

Guidance

Scores are calibrated policy signals, not verdicts. Use category_scores and set your own thresholds per category to match your risk tolerance — a stricter product may act on scores well below the default flag threshold, a lenient one only on flagged.
Moderate on both sides of a generation. Screen user input before it reaches an agent, and screen the model’s reply too. For streamed generations, moderate the assembled text after the stream completes — partial fragments classify poorly.
Batch for throughput. Send up to 32 inputs in one call; each is classified independently and returned in order.
Self-hosted. Requests run entirely on your organization’s own BRAHMA GPU. Nothing is sent to a third-party moderation service.

Moderation reuses your BRAHMA LLM plane. If that plane is still warming up, the endpoint returns a 503 with Retry-After — retry shortly.

Resources

Authentication & Errors

Voice Engine API

Bibha BRAHMA LLM

Bibha VAANI TTS

Bibha LIPI STT

Model

Backends

Categories

Request fields

Response

Example

Guidance

​Model

​Backends

​Categories

​Request fields

​Response

​Example

​Guidance

Model

Backends

Categories

Request fields

Response

Example

Guidance