django-agent-governance

AI governance and PII protection for Django — plug in, don't rip out.
Like django-allauth for AI security. Add decorators, audit logs, and a unified LLM proxy to any Django project in minutes.

Features

Install

Lightweight (regex PII engine — no extra downloads):

pip install django-agent-governance

Enterprise (Presidio NLP engine — ~500MB spaCy model):

pip install django-agent-governance[presidio]
python -m spacy download en_core_web_lg

With cloud LLM SDKs:

pip install django-agent-governance[openai,anthropic]

Quick Start

1. Add to INSTALLED_APPS

# settings.py
INSTALLED_APPS = [
    ...
    "agent_governance",
]

AGENT_GOVERNANCE = {
    "LLM_BACKEND": "ollama",            # Default: local Ollama. No API key needed.
    "OLLAMA_BASE_URL": "http://localhost:11434",
    "OLLAMA_MODEL": "llama3.2:1b",
    # Uncomment to use cloud:
    # "LLM_BACKEND": "openai",
    # "OPENAI_API_KEY": env("OPENAI_API_KEY"),
}

2. Run migrations

python manage.py migrate agent_governance

3. Use the decorators

from agent_governance.decorators import redact_pii, audit_llm_call
from agent_governance.proxy import LLMProxy

proxy = LLMProxy()

@redact_pii
@audit_llm_call
def chat_view(request):
    # request body is already PII-free ✅
    prompt = request.POST.get("message", "")
    response = proxy.chat(prompt)
    return JsonResponse({"reply": response.content})

4. View the Admin dashboard

Visit /admin/agent_governance/ to see your LLM cost tracker, PII hit rates, and response quality.

Configuration Reference

AGENT_GOVERNANCE = {
    # ── LLM Backend ───────────────────────────────────────────────
    "LLM_BACKEND": "ollama",       # "ollama" | "openai" | "anthropic" | "gemini"

    # ── Ollama (default — local, free) ────────────────────────────
    "OLLAMA_BASE_URL": "http://localhost:11434",
    "OLLAMA_MODEL": "llama3.2:1b",

    # ── OpenAI ────────────────────────────────────────────────────
    "OPENAI_API_KEY": "",          # or set env var OPENAI_API_KEY
    "OPENAI_MODEL": "gpt-4o-mini",

    # ── Anthropic ─────────────────────────────────────────────────
    "ANTHROPIC_API_KEY": "",       # or set env var ANTHROPIC_API_KEY
    "ANTHROPIC_MODEL": "claude-3-haiku-20240307",

    # ── Google Gemini ─────────────────────────────────────────────
    "GEMINI_API_KEY": "",          # or set env var GEMINI_API_KEY
    "GEMINI_MODEL": "gemini-1.5-flash",

    # ── PII Engine ────────────────────────────────────────────────
    "PII_ENGINE": "regex",         # "regex" (default) | "presidio" (install [presidio] extra)
    "PII_ENTITIES": None,          # None = all; or ["EMAIL_ADDRESS", "PHONE_NUMBER"]

    # ── Governance ────────────────────────────────────────────────
    "TOKEN_BUDGET_ENABLED": False,
    "AUDIT_LOG_RETENTION_DAYS": 90,
}

LLM-Agnostic Proxy

Switch providers without changing a line of your application code:

from agent_governance.proxy import LLMProxy

proxy = LLMProxy()

# Talks to Ollama by default
response = proxy.chat("Summarize this document.")

# Override per-call
response = proxy.chat("Summarize this.", backend="anthropic", model="claude-3-opus-20240229")

print(response.content)         # The completion text
print(response.input_tokens)    # Token counts
print(response.latency_ms)      # Latency

Supported PII Entities (Regex Engine)

Upgrade to [presidio] for: PERSON, ORG, LOCATION, DATE_TIME, NRP, and 50+ more.

Contributing

See CONTRIBUTING.md. PRs welcome!

License

MIT