PII/PHI Guardrails[NEW]

Overview

PII/PHI Guardrails apply automated PII detection and masking to the EKG ingestion pipeline before vectorization, helping ensure that sensitive data is not embedded in raw form. This helps keep raw PII/PHI out of the vector database and supports HIPAA and GDPR compliance efforts.

To protect agent inputs and outputs at runtime, see Agent PII/PHI Guardrails.

Core Concepts

Pre-embedding Sanitization

Sanitization is applied after chunking and before vectorization. Masking is applied to both the source document view and the extracted chunks, and also to document metadata attributes.

Compliance Presets

One-click compliance presets (HIPAA Safe Harbor, GDPR) automatically target the relevant entities.

Detection Methods

Regex, Deep Learning, and LLM-based detection, with configurable combinations.

Masking Strategies

Strategy	Description
Full Masking	Replaces the entire detected value
Partial Masking	Masks part of the detected value
Tokenization	Replaces the value with a token, maintaining data utility for LLM reasoning while protecting the sensitive value

Audit Log

An immutable audit log captures all detection and masking actions, creating an audit trail for regulatory review and compliance reporting.

Human Review Workflow for PII/PHI

A configurable human-in-the-loop review process lets reviewers validate guardrail detections before sensitive data is embedded.

Review Workflows

Workflow	Behavior
Full Review	Manual inspection of every detection
Confidence-Gated	High-confidence detections are masked automatically; uncertain detections are flagged for review
Auto-Approve	Zero-touch; no manual review

How Review Works

Files that require review are halted at Review Guardrails status. Embedding is blocked until the review is resolved.
Open the Chunk Viewer to review the detections. You can:
- Approve or remove detections
- Fix false positives
- Manually assign masks to false negatives
When the review is complete, the document is released for embedding automatically.

Configurable TTL expiry actions and the auto-approve fallback prevent files from being held in review indefinitely.

Benefits

Provides human oversight through chunk-level review before embedding is permitted
Reduces false positive risk through confidence-gated automation with manual override
Supports regulatory compliance by maintaining a clear chain of custody for all PII handling decisions