Skip to main content
Dome’s configuration system lets you precisely control which guards run, how they execute, and what detectors they use. This guide covers all configuration options.

​
Configuration Hierarchy

Dome organizes protection in three levels: 
Guardrail (input/output)
    └── Guard (security, moderation, privacy)
            └── Detector (specific detection method)
Each level has its own configuration options that can be customized.

​
Guardrail Configuration

​
Basic Structure

config = {
    # List of guards for each guardrail
    "input-guards": ["security-guard", "moderation-guard"],
    "output-guards": ["privacy-guard", "moderation-guard"],

    # Guardrail-level execution settings
    "input-early-exit": True,
    "input-run-parallel": False,
    "output-early-exit": True,
    "output-run-parallel": False,

    # Guard definitions (see below)
    "security-guard": { ... },
    "moderation-guard": { ... },
    "privacy-guard": { ... }
}

​
Guardrail Options

OptionTypeDefaultDescription
input-guardsList[]Guards to run on input
output-guardsList[]Guards to run on output
input-early-exitBooleanTrueStop on first input flag
input-run-parallelBooleanFalseRun input guards in parallel
output-early-exitBooleanTrueStop on first output flag
output-run-parallelBooleanFalseRun output guards in parallel

​
Execution Modes

Early Exit (default): Stops processing when the first guard flags content. Faster for rejecting clearly malicious input. 
config = {
    "input-early-exit": True,  # Stop at first flag
    "input-guards": ["security-guard", "moderation-guard"]
}
# If security-guard flags, moderation-guard won't run
Complete Execution: Runs all guards regardless of flags. Useful for comprehensive logging. 
config = {
    "input-early-exit": False,  # Run all guards
    "input-guards": ["security-guard", "moderation-guard"]
}
# Both guards always run, all flags recorded
Parallel Execution: Runs guards simultaneously for lower latency. 
config = {
    "input-run-parallel": True,
    "input-early-exit": False  # Often paired with parallel
}

​
Guard Configuration

​
Structure

Each guard groups detectors of the same type: 
config = {
    "input-guards": ["my-security-guard"],

    "my-security-guard": {
        "type": "security",                    # Required: guard type
        "methods": ["prompt-injection-mbert"], # Required: detectors
        "early-exit": True,                    # Optional
        "run-parallel": False,                 # Optional
        "blocked-response": "Request blocked"  # Optional
    }
}

​
Guard Types

TypeUse CaseAvailable Detectors
securityAdversarial attacksPrompt injection, encoding detection
moderationHarmful contentToxicity, profanity, hate speech
privacySensitive dataPII detection, secrets
integrityData qualityFormat validation (experimental)
genericCustom logicUser-defined detectors

​
Guard Options

OptionTypeDefaultDescription
typeStringRequiredGuard category
methodsListRequiredDetectors to use
early-exitBooleanTrueStop on first detector flag
run-parallelBooleanFalseRun detectors in parallel
blocked-responseStringDefaultCustom block message

​
Detector Configuration

​
Available Detectors

Security Detectors:
DetectorDescriptionOptions
prompt-injection-mbertMultilingual BERT modelthreshold
prompt-injection-deberta-v3-baseDeBERTa v3 modelthreshold
encoding-heuristicsBase64, Unicode tricksNone
security-embeddingsSemantic similaritythreshold, top_k
security-llmLLM-based detectionmodel_name
Moderation Detectors:
DetectorDescriptionOptions
moderation-flashtextFast keyword matchingwordlist
moderation-debertaNeural toxicity classifierthreshold
moderations-oai-apiOpenAI Moderation APINone
moderation-llamaguardLlama Guard modelthreshold
Privacy Detectors:
DetectorDescriptionOptions
privacy-presidioPII entity recognitionentities, threshold
detect-secretsCredential detectionNone

​
Detector-Level Configuration

Configure individual detectors within a guard: 
config = {
    "input-guards": ["security-guard"],

    "security-guard": {
        "type": "security",
        "methods": ["prompt-injection-mbert", "security-llm"],

        # Detector-specific settings
        "prompt-injection-mbert": {
            "threshold": 0.8  # Confidence threshold
        },
        "security-llm": {
            "model_name": "gpt-4o"  # Model to use
        }
    }
}

​
Common Detector Options

Threshold: Confidence score required to flag (0.0-1.0) 
"prompt-injection-mbert": {
    "threshold": 0.9  # Higher = fewer false positives
}
Model Selection: For LLM-based detectors 
"security-llm": {
    "model_name": "gpt-4o-mini"  # Faster, cheaper
}

​
TOML Configuration

Store configuration in a file for easier management: 
# dome-config.toml
[guardrail]
input-guards = ["security-guard", "moderation-guard"]
output-guards = ["privacy-guard"]
input-early-exit = true
input-run-parallel = false

[security-guard]
type = "security"
methods = ["prompt-injection-deberta-v3-base", "encoding-heuristics"]
early-exit = true

[security-guard.prompt-injection-deberta-v3-base]
threshold = 0.85

[moderation-guard]
type = "moderation"
methods = ["moderation-flashtext", "moderation-deberta"]

[privacy-guard]
type = "privacy"
methods = ["privacy-presidio"]

[privacy-guard.privacy-presidio]
entities = ["PERSON", "EMAIL", "PHONE_NUMBER", "CREDIT_CARD"]
Load from file: 
from vijil_dome import Dome

dome = Dome("dome-config.toml")
Use lowercase true and false for booleans in TOML files.

​
Configuration Examples

​
Minimal Security

Fast, low-latency protection: 
config = {
    "input-guards": ["security-guard"],
    "input-early-exit": True,

    "security-guard": {
        "type": "security",
        "methods": ["prompt-injection-mbert"]
    }
}

​
Comprehensive Protection

Full coverage for sensitive applications: 
config = {
    "input-guards": ["security-guard", "moderation-guard"],
    "output-guards": ["privacy-guard", "moderation-guard"],
    "input-early-exit": False,
    "output-early-exit": False,

    "security-guard": {
        "type": "security",
        "methods": [
            "prompt-injection-deberta-v3-base",
            "encoding-heuristics",
            "security-embeddings"
        ],
        "run-parallel": True
    },
    "moderation-guard": {
        "type": "moderation",
        "methods": ["moderation-deberta", "moderations-oai-api"],
        "run-parallel": True
    },
    "privacy-guard": {
        "type": "privacy",
        "methods": ["privacy-presidio", "detect-secrets"]
    }
}

​
Privacy-Focused

For healthcare, finance, or regulated industries: 
config = {
    "input-guards": ["security-guard"],
    "output-guards": ["privacy-guard"],

    "security-guard": {
        "type": "security",
        "methods": ["prompt-injection-deberta-v3-base"]
    },
    "privacy-guard": {
        "type": "privacy",
        "methods": ["privacy-presidio"],
        "privacy-presidio": {
            "entities": [
                "PERSON", "EMAIL", "PHONE_NUMBER",
                "CREDIT_CARD", "US_SSN", "MEDICAL_LICENSE"
            ],
            "threshold": 0.7
        }
    }
}

​
Low-Latency Production

Optimized for speed: 
config = {
    "input-guards": ["fast-security"],
    "output-guards": ["fast-moderation"],
    "input-early-exit": True,
    "output-early-exit": True,

    "fast-security": {
        "type": "security",
        "methods": ["prompt-injection-mbert"],  # Fastest model
        "early-exit": True
    },
    "fast-moderation": {
        "type": "moderation",
        "methods": ["moderation-flashtext"],  # Keyword-based, very fast
        "early-exit": True
    }
}

​
Loading Configuration from Console

Pull configuration from your Vijil Console setup: 
import os
from vijil_dome import Dome

dome = Dome.create_from_vijil_agent(
    agent_id="your-agent-id",
    api_key=os.environ["VIJIL_API_KEY"]
)
This keeps your code and configuration in sync across environments.

​
Next Steps

Using Guardrails

Runtime integration patterns

Custom Detectors

Build your own detectors

Observability

Monitoring and tracing

Framework Guides

Framework-specific integration
Last modified on March 19, 2026