llm-guardrails-kit 0.1.0

A compliance control layer between your app and any LLM: PII/PCI redaction, cost tracking, and privacy-safe audit logging. GCC-focused.

llm-guardrails-kit

A compliance control layer that sits between your application and any LLM API — redacting PII/PCI, tracking cost, and writing privacy-safe audit logs.

Built by Gulf Business Machines (GBM) with a focus on regulated GCC sectors — finance, healthcare, and government.

---

Why this exists

Sending raw customer data — names, card numbers, Emirates IDs, IBANs — to a third-party LLM provider is a data-protection and compliance risk for organizations in regulated sectors. Once data leaves your perimeter you lose control of where it is stored, logged, or used for training.

llm-guardrails-kit gives you a thin, auditable control layer you own:

• 🛡️ Redaction — detect and mask PII/PCI before the prompt leaves your network, and transparently restore the originals in the response.
• 💰 Cost governance — meter every request and roll up spend per model.
• 📝 Auditability — append a structured, privacy-safe JSON-Lines record for every call, proving what was sent without ever recording the raw values.

It is provider-agnostic (works with any LLM via a callable you supply) and the core install has zero third-party dependencies — easy to vet, easy to run in locked-down environments.

Data-protection regimes across the GCC (for example the UAE PDPL, DIFC and ADGM data-protection regulations, Saudi PDPL, and sector rules from regulators such as central banks and health authorities) place obligations on how personal data is processed and shared. This library is a control to help you meet those obligations — see the disclaimer.

---

Install

# Core — zero third-party dependencies
pip install llm-guardrails-kit

# Optional: LiteLLM integration (one client for 100+ providers)
pip install "llm-guardrails-kit[litellm]"

# Development (pytest, build, ruff)
pip install "llm-guardrails-kit[dev]"

Requires Python 3.11+.

---

Quickstart

Redact a prompt containing a card number, Emirates ID, and email; send it through a GuardedClient with a dummy completion function; and inspect the audit record — all without an API key.

from llm_guardrails_kit import GuardedClient, AuditLogger, CostTracker

# 1. A stand-in for your real LLM call. It receives the ALREADY-REDACTED prompt.
#    Swap this for OpenAI, Anthropic, LiteLLM, Bedrock — anything.
def my_llm(prompt: str, model: str, **kwargs) -> dict:
    # A real LLM never sees the raw PII — only placeholders like [CARD_1].
    print("LLM received:", prompt)
    return {
        "text": "Thanks! I've noted the details for [EMAIL_1].",
        "input_tokens": 42,
        "output_tokens": 12,
    }

# 2. Wire up the control layer.
audit = AuditLogger(path="audit.jsonl")
client = GuardedClient(my_llm, cost_tracker=CostTracker(), audit_logger=audit)

# 3. Call it with sensitive data.
result = client.complete(
    "Charge card 5555 5555 5555 4444 for Emirates ID 784-1984-1234567-4 "
    "and email the receipt to [e-mail]",
    model="gpt-4o",
)

print("Returned to app:", result.text)
print("Redaction counts:", result.redaction_counts)
print("Cost (USD):", result.cost.total_cost)
audit.close()

Output (abridged):

LLM received: Charge card [CARD_1] for Emirates ID [EMIRATES_ID_1] and email the receipt to [EMAIL_1]
Returned to app: Thanks! I've noted the details for [e-mail].
Redaction counts: {'credit_card': 1, 'emirates_id': 1, 'email': 1}
Cost (USD): 0.000225

The audit.jsonl record — note there is no raw PII, only counts:

{"input_tokens": 42, "latency_ms": 0.05, "metadata": {}, "model": "gpt-4o", "output_tokens": 12, "redaction_counts": {"credit_card": 1, "email": 1, "emirates_id": 1}, "request_id": "d3b0...", "timestamp": "2026-07-03T09:15:22.104531Z", "total_cost": 0.000225}

Redaction on its own

from llm_guardrails_kit import Redactor

r = Redactor()
res = r.redact("Wire to AE07 0331 2345 6789 0123 456, call [phone]")
print(res.redacted_text)  # Wire to [IBAN_1], call [PHONE_1]
print(res.counts)         # {'iban': 1, 'phone': 1}

# Fully reversible for the caller:
print(r.restore("Confirmed [IBAN_1]", res.mapping))  # Confirmed AE07 0331 2345 6789 0123 456

Partial masking (show only the last 4 of a card)

from llm_guardrails_kit import Redactor, RedactionPolicy, MaskStyle

policy = RedactionPolicy(mask_style=MaskStyle.PARTIAL)
print(Redactor(policy).redact("Card 5555 5555 5555 4444").redacted_text)
# Card **** **** **** 4444

Using it with LiteLLM (optional extra)

from llm_guardrails_kit import GuardedClient, AuditLogger

client = GuardedClient.from_litellm(audit_logger=AuditLogger(path="audit.jsonl"))
result = client.complete("Email [e-mail]", model="gpt-4o")

---

What gets detected

Detector	Placeholder	Validation
Email	[EMAIL_n]	RFC-ish pattern
Phone (intl + UAE/GCC)	[PHONE_n]	E.164 digit-count (7–15)
Credit card	[CARD_n]	Luhn checksum + length 13–19
Generic PCI	[PCI_n]	4×4 card-shaped grouping (defensive)
IBAN	[IBAN_n]	ISO 13616 mod-97 + per-country length
Emirates ID	[EMIRATES_ID_n]	784-YYYY-NNNNNNN-C format + Luhn
IPv4 address	[IP_n]	Octet range 0–255
Custom	[NAME_n]	Your regex

Every numeric detector uses real structural validation, not just a regex, so 1234 5678 9012 3456 is not mistaken for a valid card, and a corrupted IBAN is rejected by its checksum.

---

API overview

Symbol	Purpose
Redactor(policy=None)	.redact(text) -> RedactionResult; .restore(text, mapping) -> str
RedactionPolicy	Toggle detectors, add custom_patterns, pick mask_style
RedactionResult	.redacted_text, .mapping, .counts, .total_redactions
MaskStyle	PLACEHOLDER (default, reversible), PARTIAL (last 4), FULL
CostTracker(prices=None, default_price=None)	.record(model, in, out), .total(), .reset(), .set_price()
DEFAULT_PRICES	Editable per-1M-token price table (verify before production!)
AuditLogger(path=None, sink=None)	.log(...) -> AuditRecord; JSON-Lines, never logs raw PII
GuardedClient(completion_fn, ...)	.complete(prompt, model); .from_litellm(...) optional helper
luhn_valid, iban_valid, emirates_id_valid	Standalone validators

---

Cost prices — verify before you trust them

DEFAULT_PRICES in cost.py ships a small, illustrative table of per-1M-token prices for a few common OpenAI and Anthropic models. Prices change frequently — always confirm against your provider's current pricing and override:

from llm_guardrails_kit import CostTracker, ModelPrice

tracker = CostTracker()
tracker.set_price("gpt-4o", ModelPrice(input_per_1m=2.5, output_per_1m=10.0))

---

Development

pip install -e ".[dev]"
ruff check .
pytest
python -m build

---

Disclaimer

llm-guardrails-kit is a helper control layer, not a guarantee of regulatory compliance. It reduces the risk of leaking sensitive data to third-party LLMs, but:

• No automated redactor detects 100% of sensitive data in free-form text.
• It does not claim any specific certification or regulatory approval.
• You remain responsible for validating it against your own legal, regulatory, and contractual obligations before relying on it in production.

Test it thoroughly with your own data and threat model. Use of this software is subject to the MIT License and provided "as is", without warranty.

---

Built by Gulf Business Machines (GBM). Maintainer: Hasan Odeh.