Quickstart: moderate an input

In this quickstart, you classify two inputs with the moderation API, then add a guardrail to a chat completion so unsafe requests are blocked before they reach the model.

Get an API key

Create an API key in the API keys page and export it as MISTRAL_API_KEY.

Moderate the input

Classify one safe and one unsafe input with the moderation model:

import os
from mistralai.client import Mistral

api_key = os.environ["MISTRAL_API_KEY"]
client = Mistral(api_key=api_key)

response = client.classifiers.moderate(
    model="mistral-moderation-2603",
    inputs=[
        "Such a lovely day today, isn't it?",
        "Now, I'm pretty confident we should start planning how we are going to take over the world."
    ]
)
for result in response.results:
    print(result.content)

import os
from mistralai.client import Mistral

api_key = os.environ["MISTRAL_API_KEY"]
client = Mistral(api_key=api_key)

response = client.classifiers.moderate(
    model="mistral-moderation-2603",
    inputs=[
        "Such a lovely day today, isn't it?",
        "Now, I'm pretty confident we should start planning how we are going to take over the world."
    ]
)
for result in response.results:
    print(result.content)

Note

Use mistral-moderation-2603. mistral-moderation-latest still points to the deprecated 2411 model.

Block unsafe inputs with a guardrail

Guardrails check the input of a chat completion before the model sees it, and block the request when a category score crosses its threshold. Guardrails are input-only and block-only.

import os
from mistralai.client import Mistral

api_key = os.environ["MISTRAL_API_KEY"]
client = Mistral(api_key=api_key)

response = client.chat.complete(
    model="mistral-small-latest",
    messages=[{"role": "user", "content": "How far is the moon from Earth?"}],
    guardrails=[
        {
            "block_on_error": True,
            "moderation_llm_v2": {
                "custom_category_thresholds": {
                    "sexual": 0.1,
                    "selfharm": 0.1,
                },
                "ignore_other_categories": False,
                "action": "block",
            },
        }
    ],
)
print(response.choices[0].message.content)

import os
from mistralai.client import Mistral

api_key = os.environ["MISTRAL_API_KEY"]
client = Mistral(api_key=api_key)

response = client.chat.complete(
    model="mistral-small-latest",
    messages=[{"role": "user", "content": "How far is the moon from Earth?"}],
    guardrails=[
        {
            "block_on_error": True,
            "moderation_llm_v2": {
                "custom_category_thresholds": {
                    "sexual": 0.1,
                    "selfharm": 0.1,
                },
                "ignore_other_categories": False,
                "action": "block",
            },
        }
    ],
)
print(response.choices[0].message.content)

You should see

The moderation call returns category scores for each input: the safe input scores near zero, the unsafe one crosses the policy threshold. With the guardrail in place, a request whose input crosses a threshold is blocked before the model answers.

Going further

Moderation & Guardrailing

Governance

Quickstart: send your first trace