Guardrails

Guardrails is the surface where you configure the three filter layers Tale runs on every chat message in your organisation. Each message passes through content safety (word lists and admin regex), then PII detection (built-in patterns plus custom), then an optional external moderation provider — in that fixed order, on the way in and on the way out. Admins and Owners read this page when a regulator names a content rule, when a leak warrants a tighter policy, or when an agent's replies need to be sanitised before they leave the model.

A worked layering

To configure the layers, open Settings > Governance > Guardrails. The overview shows three status cards, one per layer — content safety, PII detection, moderation. Each card links to its own configuration page where you pick whether the layer runs on input, on output, or both, and what it does on a match (block the message, mask the match, or flag and pass). The recent-events table at the bottom of the overview shows the last 50 detections, blocks, and provider errors with their layer, direction, and match category.

Content safety

Content safety is the layer you own. Define one or more categories — hate speech, profanity, a custom regex for an internal codename — and pick a mode per category: block refuses the message, mask replaces matches with a placeholder, flag records the detection without changing the message. Block wins over mask wins over flag when more than one category matches.

The layer's word lists and patterns never leave the deployment. Matched text is not stored — only the category, the direction (input or output), and the count of matches end up in the audit event.

PII detection

PII detection ships with patterns for emails, phones, government IDs, payment numbers, and a long tail of regional formats. Add custom patterns if your regulator names a format the built-ins miss. Pick a mode — block, mask with a placeholder, or flag — and an apply direction. Mask is the typical choice for output filtering when the model has been given access to records that include PII it should not echo back.

Moderation provider

The moderation layer is an external classifier — OpenAI Moderation, Azure Content Safety, Perspective API, or a custom HTTP endpoint. Configure the provider's endpoint, an API key, and the category-to-action mapping (each provider returns its own taxonomy; the mapping decides which categories block, mask, or flag). The layer is optional — leave it disabled and only the first two layers run.

The provider sits on the network egress path. Failures are configurable per direction: fail-open lets the message through, fail-closed refuses it. The recent-events view shows provider errors, HTTP statuses, and circuit-open events when the layer is rate-limited.

Recent events

Every detection, block, and provider error lands in the recent-events table for 30 days. Filter by layer or by kind; click a row to see the matched categories, the actor, the message id, and the timestamp. Raw matched text is never stored — the events are a tuning surface, not a content archive.

Where this fits

Guardrails is the runtime filter between the user and the model in both directions. Pair it with content and models so an approved model is also subject to the approved content rules. The companion is the audit log — every block and every mask the guardrail layers apply lands there as a permanent record.