Implementation:Protectai Llm guard Vault

Knowledge Sources	LLM Guard
Domains	Security, Data_Privacy, State_Management
Last Updated	2026-02-14 12:00 GMT

Overview

Concrete tool for storing anonymization placeholder-to-original-value mappings provided by the LLM Guard library.

Description

The Vault class is an in-memory tuple store that holds (placeholder, original_value) pairs created during anonymization. It is shared between the Anonymize input scanner (which populates it) and the Deanonymize output scanner (which reads from it). The class provides methods to append, extend, remove, and query stored tuples.

Usage

Import and instantiate this class before creating Anonymize and Deanonymize scanners. Pass the same Vault instance to both scanners to enable reversible anonymization across the LLM call boundary.

Code Reference

Source Location

• Repository: llm-guard
• File: llm_guard/vault.py
• Lines: L4-35

Signature

class Vault:
    """
    Stores tuples of (placeholder, original_value) for anonymization/deanonymization.
    """

    def __init__(self, tuples: Optional[List[Tuple]] = None):
        """
        Args:
            tuples: Pre-existing placeholder-to-original mappings. Default creates empty vault.
        """

    def append(self, new_tuple: Tuple) -> None:
        """Add a single (placeholder, original) mapping."""

    def extend(self, new_tuples: List[Tuple]) -> None:
        """Add multiple mappings at once."""

    def remove(self, tuple_to_remove: Tuple) -> None:
        """Remove a specific mapping."""

    def get(self) -> List[Tuple]:
        """Return all stored mappings."""

    def placeholder_exists(self, placeholder: str) -> bool:
        """Check if a placeholder already exists in the vault."""

Import

from llm_guard.vault import Vault

I/O Contract

Inputs

Name	Type	Required	Description
tuples	Optional[List[Tuple]]	No	Pre-existing placeholder mappings (default: None creates empty vault)

Outputs

Name	Type	Description
Vault instance	Vault	Object with append, extend, remove, get, placeholder_exists methods
get()	List[Tuple]	All stored (placeholder, original_value) tuples
placeholder_exists()	bool	Whether a given placeholder string exists in the vault

Usage Examples

Basic Vault Usage

from llm_guard.vault import Vault

# 1. Create empty vault
vault = Vault()

# 2. Pass to both Anonymize and Deanonymize
from llm_guard.input_scanners import Anonymize
from llm_guard.output_scanners import Deanonymize

anonymize = Anonymize(vault)
deanonymize = Deanonymize(vault)

# 3. After scanning prompt, vault is populated
sanitized_prompt, is_valid, score = anonymize.scan("My name is John Smith")
print(vault.get())  # [("[REDACTED_PERSON_1]", "John Smith")]

# 4. Deanonymize reads from the same vault
output = "Hello [REDACTED_PERSON_1], your request is processed."
restored, _, _ = deanonymize.scan(sanitized_prompt, output)
print(restored)  # "Hello John Smith, your request is processed."

Pre-populated Vault

from llm_guard.vault import Vault

# Initialize with existing mappings
vault = Vault(tuples=[
    ("[REDACTED_PERSON_1]", "Jane Doe"),
    ("[REDACTED_EMAIL_ADDRESS_1]", "jane@example.com"),
])

# Check for existing placeholders
vault.placeholder_exists("[REDACTED_PERSON_1]")  # True
vault.placeholder_exists("[REDACTED_PERSON_2]")  # False

Related Pages

Implements Principle

• Principle:Protectai_Llm_guard_Vault_State_Management

Requires Environment

• Environment:Protectai_Llm_guard_Python_Runtime_Dependencies