Moderation

Run batch content moderation through OpenAI omni-moderation and Mistral with Batchwork's normalized flagged/category verdicts.

Moderation batches are supported through OpenAI (omni-moderation-latest, text plus image_urls) and Mistral (mistral-moderation-latest, text only). Pass the model as a provider/model string.

CLI

Use batchwork submit moderations to return after registration or batchwork run moderations for the complete lifecycle. JSON, JSONL, CSV, and text sources map to BatchModerationRequest; each request needs a value, image_urls, or both.

batchwork --json run moderations values.txt \
  --model openai/omni-moderation-latest

CSV columns are custom_id, value, and image_url (at least one of value/image_url per row). Text input supplies one value per line. There are no modality default flags.

OpenAI

from batchwork import BatchModerationRequest, Batchwork

async with Batchwork() as client:
    job = await client.batch_moderations(
        model="openai/omni-moderation-latest",
        requests=[
            BatchModerationRequest(custom_id="comment-1", value="User comment text."),
            BatchModerationRequest(
                custom_id="upload-1",
                image_urls=["https://example.com/upload.png"],
            ),
        ],
    )
    await job.wait(timeout=3600)
    verdict = (await job.collect())[0].moderation

OpenAI text-only requests serialize input as the string; requests with image_urls serialize input as a parts list mixing text and image_url entries.

Mistral

job = await client.batch_moderations(
    model="mistral/mistral-moderation-latest",
    requests=[BatchModerationRequest(custom_id="comment-1", value="User comment text.")],
)

Mistral moderation is text-only: image_urls are rejected locally before upload. Mistral omits a top-level flagged field, so Batchwork computes it as “any category flagged”.

Result shape

result.moderation  # BatchModeration | None

A BatchModeration contains flagged, categories (dict[str, bool]), and category_scores (dict[str, float]) with provider-native category names.

Unsupported providers

Anthropic, Azure, Google Gemini, Groq, Together AI, and xAI moderation submissions fail locally.

See Provider overview and Results.