Content Policy

The content policy guardrail enforces your organization's acceptable use policies and can catch AI hallucinations by cross-referencing email content against a knowledge base you provide.

Default behavior

Out of the box, Sluice flags emails that contain:

• Spam or unsolicited commercial content
• Phishing attempts or social engineering
• Illegal content or content that violates common organizational standards

Custom content policies

Replace the default policy with your own definition in Settings > Guardrails > Content Policy (up to 2,000 characters).

Examples:

• "Never mention competitor products by name. Never offer discounts or refunds without directing the customer to our support portal."

• "All product claims must be verifiable. Do not promise features that are not listed in our official documentation."

• "Emails must not include internal jargon, project codenames, or references to internal tools or processes."

Hallucination detection with Knowledge Base

This is one of Sluice's most powerful features. Upload your product documentation, pricing, policies, or FAQ as a knowledge base, and Sluice will cross-check AI-generated claims against your source of truth.

How it works

1. Go to Settings > Content Policy > Knowledge Base
2. Use the rich text editor to enter your reference material (up to 64 KB)
3. When an email is analyzed, the content policy guardrail checks AI-generated claims against your knowledge base
4. Contradictions are flagged

Strict mode

Example: Your knowledge base says your Enterprise plan supports up to 50 users. An AI agent writes: "Our Enterprise plan includes unlimited users."

• With strict mode off: Flagged — this directly contradicts the knowledge base
• With strict mode on: Also flagged — any claim about plan limits is checked against the knowledge base

With strict mode on, if the agent writes "Our Enterprise plan includes 24/7 phone support" and the knowledge base doesn't mention phone support, this would also be flagged — the claim is within the domain of plan features but isn't covered.

Best practices for your knowledge base

• Include facts your agents are likely to reference — pricing, plan features, policies, return windows, SLAs
• Keep it up to date — stale entries cause false positives when your product changes
• Use strict mode for high-stakes communication — healthcare, finance, legal
• Use non-strict mode for general support — catches outright errors but gives agents flexibility to address topics not explicitly in the knowledge base

Use cases

SaaS support — Upload your pricing page and feature matrix. Catch agents that invent features, misquote prices, or promise capabilities that don't exist.

E-commerce — Upload your return policy and warranty terms. Catch agents that offer unauthorized refunds or misstate warranty coverage.

Healthcare — Upload approved patient communication templates. Use strict mode to ensure agents never make medical claims outside approved language.