coreason-aegis 0.3.1

bidirectional privacy filter that sits between the CoReason platform and the external world

coreason-aegis

A bidirectional privacy filter that sits between the CoReason platform and the external world.

coreason-aegis implements a "Tokenize & Detokenize" strategy. Before any text leaves the secure perimeter (e.g., to OpenAI), Aegis scans it for sensitive entities (Names, MRNs, SSNs, Emails) and replaces them with consistent, context-aware tokens. When the LLM responds using these tokens, Aegis intercepts the message and re-injects the real data only for the authorized user's eyes.

Features

• Scan-Mask-Map-Reveal Loop: Deterministic tokenization ensures LLM reasoning consistency.
• The "Vault" of Identity: Ephemeral, encrypted mapping tables.
• Context Preservation: Uses tokens like [PATIENT_A] instead of [REDACTED].
• Custom Recognizers: Supports custom entities like MRNs, Protocol IDs, and Lot Numbers.

For detailed requirements, see Product Requirements.

Installation

pip install coreason-aegis

Or using Poetry:

poetry add coreason-aegis

Note: You may need to download the Spacy model manually if not handled by the package installer:

python -m spacy download en_core_web_lg

Usage

from coreason_aegis.main import Aegis
from coreason_aegis.models import AegisPolicy, RedactionMode

# Initialize Aegis
aegis = Aegis()

# Define a policy
policy = AegisPolicy(
    allow_list=["Tylenol"],
    entity_types=["PERSON", "EMAIL", "PHONE_NUMBER"],
    mode=RedactionMode.REPLACE,
    confidence_score=0.7
)

# Sanitize user prompt
user_prompt = "Patient John Doe (DOB: 12/01/1980) has a rash."
session_id = "session_123"
sanitized_prompt, mapping_id = aegis.sanitize(user_prompt, policy, session_id)

print(f"Sanitized: {sanitized_prompt}")
# Output: "Patient [PATIENT_A] (DOB: [DATE_B]) has a rash."

# ... Send to LLM ...
llm_response = "For [PATIENT_A], considering the rash..."

# Desanitize LLM response
final_response = aegis.desanitize(llm_response, session_id, mapping_id)

print(f"Final: {final_response}")
# Output: "For John Doe, considering the rash..."