Privacy Sanitization

Use privacy-sanitization to sanitize secrets before /v1/chat/completions reaches the model. Use redact for readable masked text, tokenize for stable per-request placeholders, and block to stop the request before execution. For broader security and data-handling policy, see Security.

Send the minimum request

{
  "model": "qwen/qwen3-235b-a22b-instruct-2507-fp8",
  "messages": [
    {
      "role": "user",
      "content": "Summarize this incident note, but hide apiKey=sk-live-demo-123456 and DATABASE_URL=postgres://app:super-secret@db.internal:5432/gonkagate."
    }
  ],
  "plugins": [
    {
      "id": "privacy-sanitization",
      "mode": "redact"
    }
  ]
}

Expected result: the request continues with sanitized input, and covered fragments are not forwarded to the model in raw form.

Choose the mode

Compare one input across all three modes

Placeholder strings below are examples only. Exact replacement strings are implementation-defined.

Input

Compare sk-live-prod-123 and sk-live-staging-456. In the next step, use sk-live-prod-123 again.

redact

Compare [SECRET:API_KEY] and [SECRET:API_KEY]. In the next step, use [SECRET:API_KEY] again.

tokenize

Compare [TOKEN_1] and [TOKEN_2]. In the next step, use [TOKEN_1] again.

block

Request rejected before model execution.

• Choose redact when the prompt should stay readable after sanitization.
• Choose tokenize when one request needs stable hidden placeholders for repeated secret references.
• Choose block when continuing with sanitized content would still be unsafe.

redact keeps the request readable but collapses secret identity. tokenize keeps the value hidden while preserving which secret is which inside the same sanitized request.

Use tokenize to preserve hidden-value identity inside one sanitized request, not to format or restore chat history.

Know the limits before rollout

• mode is required.
• Public privacy-sanitization on default /v1/chat/completions is stateless: sanitization runs on request ingress and does not keep privacy state across requests.
• redact, tokenize, and block work in both streaming and non-streaming requests.
• Do not expect _gonka.privacy.history_writeback, event: gonka.input_patch, raw restore in tool_calls, or any other client-visible history restore path.
• Do not rely on history or session state from this default public surface.
• This public managed profile is secret-oriented. It does not promise broad generic PII redaction by default.
• privacy-sanitization is currently incompatible with built-in web search, including plugins: [{ "id": "web" }] and model: "...:online". Choose one path per request.
• Authenticated requests can also inherit saved plugin settings when plugin settings policy is enabled.

Enable or disable the plugin

Request-level opt-in works when:

• plugins includes id="privacy-sanitization"
• mode is redact, tokenize, or block
• that plugin entry has enabled !== false

{
  "plugins": [
    {
      "id": "privacy-sanitization",
      "mode": "redact"
    }
  ]
}

Authenticated requests can also activate privacy-sanitization from saved plugin settings. Treat the request payload as the request-level contract, but not the only activation source when plugin settings policy is enabled.

To disable request-level activation for one request:

1. Omit privacy-sanitization from plugins if you do not use saved plugin defaults.
2. If saved plugin settings may auto-activate it, send enabled: false or disable the saved default.
3. If saved plugin settings are locked by policy, request-level overrides may be rejected.

Use it from SDKs

The OpenAI SDK request types do not model GonkaGate-specific plugins. Send the field through the runtime-specific escape hatch.

TypeScript

Use // @ts-expect-error or your own wrapper around undocumented request fields.

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.gonkagate.com/v1",
  apiKey: "gp-your-api-key"
});

const response = await client.chat.completions.create({
  model: "qwen/qwen3-235b-a22b-instruct-2507-fp8",
  messages: [
    {
      role: "user",
      content:
        "Summarize this incident note, but hide Authorization: Bearer demo-secret-token and DATABASE_URL=postgres://app:super-secret@db.internal:5432/gonkagate."
    }
  ],
  // @ts-expect-error undocumented request field for an OpenAI-compatible backend
  plugins: [{ id: "privacy-sanitization", mode: "redact" }]
});

console.log(response.choices[0]?.message?.content);

Python

Use extra_body for undocumented request fields.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.gonkagate.com/v1",
    api_key="gp-your-api-key"
)

response = client.chat.completions.create(
    model="qwen/qwen3-235b-a22b-instruct-2507-fp8",
    messages=[
        {
            "role": "user",
            "content": "Summarize this incident note, but hide apiKey=sk-live-demo-123456."
        }
    ],
    extra_body={
        "plugins": [{"id": "privacy-sanitization", "mode": "redact"}]
    }
)

print(response.choices[0].message.content)

Handle common failures

See also

• Plugins Overview to compare plugin behavior and saved defaults.
• Web Search for the separate built-in grounding path. It is currently incompatible with privacy-sanitization; choose one path per request.
• Chat Completions API reference for the exact plugin field schema.
• Security for broader security and data-handling policy.