safetype (Python)

Rule-based content moderation for Python servers: profanity, hate speech and harassment. It's the same engine and English dictionary as the SafeType.js JavaScript packages, so your backend can re-check exactly what the browser checked. It has no dependencies and runs fully offline.

pip install safetype

from safetype import create_moderation_engine

engine = create_moderation_engine(mode="mask")
engine.moderate("What the fuck are you doing?").text
# 'What the f*ck are you doing?'

Server-side enforcement

Client-side moderation can be bypassed (devtools, curl, modified clients). Re-check before you store:

engine = create_moderation_engine(
    mode="mask",
    policy={"hate": {"action": "block"}, "harassment": {"action": "block"}},
)

check = engine.validate(comment)
if not check.valid:
    raise ValidationError([v.category for v in check.violations])
store(check.result.text)                      # profanity masked
if check.result.requires_review:
    queue_for_human_review(comment)           # e.g. quoted or reported speech

FastAPI:

@app.post("/comments")
def create_comment(body: CommentIn):
    check = engine.validate(body.text)
    if not check.valid:
        raise HTTPException(422, detail=[{"category": v.category, "rule": v.rule_id} for v in check.violations])
    return save(check.result.text)

Django form:

def clean_body(self):
    check = engine.validate(self.cleaned_data["body"])
    if not check.valid:
        raise forms.ValidationError("Please keep it respectful.")
    return check.result.text

What it handles

• Five actions (allow, mask, suggest, warn, block) and per-category policies
• Profanity is masked deterministically. Hate and harassment rules are lower-confidence signals that context can weaken: quoted or reported speech becomes requires_review instead of being acted on
• Unicode (accents, full-width, Cyrillic/Greek lookalikes, zero-width), leet (sh1t), repeated letters, spaced letters
• No Scunthorpe problem, and masking is idempotent
• English and Hinglish (romanized Hindi: "bhenchod", "chutiya", "teri maa ka...") out of the box; languages=["en"] turns Hinglish off
• Custom words, custom rules, allow lists, mask styles (vowels, partial, full, or a function)

API

engine = create_moderation_engine(
    mode="mask", language="en", languages=None, categories=None, policy=None, packs=None,
    custom_words=None, allow_list=None, rules=None,
    mask_char="*", mask_style="vowels", detect_spaced_letters=True,
)
engine.moderate(text)   # ModerationResult(text, suggested_text, matches, action, allowed, requires_review, mapping)
engine.mask(text)       # str
engine.validate(text)   # ValidationResult(valid, result, violations)
engine.add_rule({"id": "app.x", "pattern": "(foo|bar*)", "category": "custom"})
engine.remove_rule("app.x")
engine.set_policy({"harassment": {"action": "block", "min_confidence": 0.7}})
engine.set_mode("warn")

safetype.moderate(text); safetype.mask(text)   # shared default engine

Offsets (match.start / match.end) are Python string indices (code points).

Parity with JavaScript

The test suite replays 684 cases recorded from the JavaScript engine and requires identical results. The dictionary is exported from the JavaScript build, so both always share the same rules.

Limitations

Rule-based detection misses novel spellings and sarcasm, and it can flag reclaimed or quoted language. Treat hate and harassment results as signals for review, not verdicts. Contextual (AI) classifiers are available in the JavaScript engine and planned for Python.

MIT licensed.