redact-prompt 0.1.0

Redact sensitive information (API keys, emails, card info, etc.) from prompts before sending to LLMs

redact-prompt

Redact PII from prompts before sending to LLMs. Restore original values in responses.

Install

pip install redact-prompt

The spaCy language model downloads automatically on first use.

How It Works

1. Redact — Regex + NER detect PII, replace with numbered placeholders
2. Inject — result.text includes instruction telling LLM to preserve placeholders
3. Restore — unredact() maps placeholders back to original values

Same values get same placeholders (deterministic), so [e-mail] appearing twice → [EMAIL_1] both times.

Quick Start

from redact_prompt import redact, unredact

result = redact("Email [e-mail] about the project")

# Send result.text to LLM (includes injected instruction to preserve placeholders)

restored = unredact(llm_response)

Examples

# OpenAI
from openai import OpenAI
from redact_prompt import redact, unredact

client = OpenAI()
result = redact("My email is [e-mail] and my API key is sk-proj-abc123xyz789def456.")

response = client.responses.create(
    model="gpt-5-mini-2025-08-07",
    input=result.text,
)
print(unredact(response.output_text))

# Anthropic
import anthropic
from redact_prompt import redact, unredact

client = anthropic.Anthropic()
result = redact("Hi, I'm John Smith from Acme Corp. My email is [e-mail].")

message = client.messages.create(
    model="claude-sonnet-4-5-20250929",
    max_tokens=1024,
    messages=[{"role": "user", "content": result.text}],
)
print(unredact(message.content[0].text))

# Google Gemini
from google import genai
from redact_prompt import redact, unredact

client = genai.Client()
result = redact("Contact me at 555-123-4567 or [e-mail] for details.")

response = client.models.generate_content(
    model="gemini-3.0-pro-preview",
    contents=result.text,
)
print(unredact(response.text))

See examples/ for OpenRouter and more.

What It Detects

Regex-based:

• Email — [e-mail] → [EMAIL_1]
• Phone — 555-123-4567 → [PHONE_1]
• SSN — 123-45-6789 → [SSN_1]
• Credit Card — 4111-1111-1111-1111 → [CREDIT_CARD_1]
• IP Address — 192.168.1.1 → [IP_1]
• API Keys — OpenAI, Anthropic, AWS, GitHub, Stripe, Slack, Google

NER-based (via spaCy):

• Person — John Smith → [PERSON_1]
• Organization — Acme Corp → [ORG_1]
• Location — New York → [LOCATION_1]

API

from redact_prompt import redact, unredact, clear

result = redact("text")        # Returns RedactionResult
result.text                    # Redacted + instruction (send to LLM)
result.redacted                # Just redacted text (for debugging)
result.entities                # List of detected entities

unredact(llm_response)         # Restore original values
clear()                        # Reset mappings between conversations

Options

# Exclude specific types
result = redact("text", exclude=["EMAIL", "PHONE"])

# Only include specific types
result = redact("text", include=["SSN", "CREDIT_CARD"])

# Disable NER (faster, regex only)
result = redact("text", use_ner=False)

# Multiple conversations (use class API)
from redact_prompt import Redactor
r = Redactor()
result = r.redact("text")
r.clear()  # Reset for new conversation

Available types: EMAIL, PHONE, SSN, CREDIT_CARD, IP, API_KEY, PERSON, ORG, LOCATION

Contributing

PRs welcome! See GitHub Issues for wanted features.

Adding a new PII pattern:

1. Add regex to PATTERNS in redact_prompt/__init__.py
2. Add test in tests/test_redact.py
3. Update "What It Detects" in README

Wanted patterns: URL, Date of Birth, Address/ZIP, Bank Account, Drivers License, Passport Number, and more. Check issues.

License

MIT