PII Detection for Insurance Claim Anonymization

Why it matters

A PII rail can protect privacy and still ruin the data. This notebook measures both sides: identity removed, and narrative signal kept for the model downstream.

What to practise

Configure rails.config.gliner with a server_endpoint, a threshold and twelve input.entities, and enable the gliner mask pii on input flow. Note why the notebook masks rather than detects: detection would refuse every claim. Run the smoke test, then the evaluation over the claim subset, and read entity-level precision, recall and over-masking rate per type. Watch which types the rail returns that the entities list never asked for, and which format variants of a phone number it misses.

How it fits

This puts 9.2.1’s prompt-side redaction into practice. The same Bite asks you to protect the other copies too — generated answers, retrieved chunks and traces — that this input rail does not cover.