Documentation hub

Start with the runtime. Add the console when the operating model calls for it.

This section is organized around adoption stages rather than product marketing. Use the runtime-first guides to get running locally, deploy in production, wire the SDKs into your app, and decide which control plane model fits your team.

Compare Plans

Start here

These pages are the shortest path from evaluation to production rollout.

Getting Started

Run the runtime locally, send a sample evaluation, and decide when to add the console.

Deploy

Move from local Docker to Kubernetes or Helm, and understand the console deployment options.

Integrate

Wrap your model calls with input, output, tool, and streaming evaluation.

Reference

Use these pages when you need deployment guidance, data-handling clarity, or an architectural view for security review and procurement.

Architecture

Runtime and control plane responsibilities, network boundary, and policy flow.

Deployment Patterns

Compare Starter, Growth, and Enterprise operating models before you buy.

Data Handling

See what runs in your infrastructure and what the hosted console is designed to retain by default.

Detector reference

Built-in detectors

The runtime ships with these detectors. Compose them into policies, mix with your own detectors, and apply on input, output, tool calls, or streaming responses.

PII

Input & Output

Detect and redact personally identifiable information across emails, phone numbers, addresses, IDs, and more.

Secrets

Input & Output

Catch leaked API keys, tokens, credentials, and other sensitive material before they reach a model or a downstream system.

Prompt injection

Input

Identify prompt-injection patterns and adversarial instruction chains aimed at overriding your application prompt.

Jailbreak

Input

Recognize jailbreak templates and evasion attempts that try to bypass policy controls.

Toxicity

Input & Output

Flag abusive, hateful, or harassing content across both user input and model output.

Profanity

Input & Output

Filter profanity with configurable severity thresholds per project and environment.

Topic control

Input & Output

Keep conversations within allowed topic boundaries or block specified off-topic areas.

Language

Input & Output

Enforce language restrictions and detect unexpected or unsupported languages in traffic.

Competitor mentions

Output

Prevent model responses from referencing competitor names or products you have excluded.

Hallucination signals

Output

Surface low-confidence or unsupported claims in responses for review or blocking.

Code safety

Output

Detect insecure code patterns, shell-escape risks, and dangerous suggestions in code responses.

URL and domain control

Input & Output

Block, allow, or inspect URLs and domains referenced inside prompts and responses.

Operating model by plan

The product boundary is simple: the runtime is always self-hosted, while the console changes by plan.

Starter

Open-source runtime only. You self-host it, manage policy files locally, and use docs-based onboarding.

Growth

Self-hosted runtime plus hosted console. The console adds policy workflow and metadata-first operational visibility.

Enterprise

Custom deployment support, including the option to self-host both runtime and console.