Semantic Configuration

The Semantic section configures Synentra's optional semantic intent analysis engine.

Unlike deterministic policy evaluation, semantic analysis uses machine learning to understand the intent behind free-form prompts and requests. This enables Synentra to identify malicious, unsafe, or out-of-scope requests even when they do not match explicit policy rules.

Semantic analysis can use either:

• the built-in ONNX/BERT model (recommended for offline deployments), or
• an external Large Language Model (LLM) such as OpenAI, Azure AI, Google Gemini, or Ollama.

When disabled, only deterministic policy evaluation is performed.

---

Properties

Property	Type	Default	Description
Enabled	bool?	false	Enables or disables semantic intent analysis.
ConfidenceThreshold	double?	0.7	Minimum confidence score (0–1) required before Synentra trusts the classification result.
AllowLowConfidence	bool?	false	Whether requests with confidence below ConfidenceThreshold are allowed.
DefaultProvider	string	"Internal"	Active provider. Supported values: Internal, OpenAi, AzureAi, Gemini, Ollama.

---

Providers

The provider selected by DefaultProvider determines which model performs semantic analysis.

{
  "Semantic": {
    "Enabled": true,
    "DefaultProvider": "Internal"
  }
}

---

Internal (ONNX)

The Internal provider performs semantic classification completely offline using a bundled ONNX BERT model.

No external network calls are made, making it suitable for air-gapped and privacy-sensitive deployments.

Properties

Property	Type	Default	Description
ModelType	string	"Community"	Model package type. Supported values: Community or Pro.
PackagePath	string?	null	Absolute path to the ONNX model package ZIP. If omitted, Synentra uses the bundled model.
LicensePath	string?	null	Absolute path to the Pro license file. Required only when ModelType is Pro.
MaxLength	int?	128	Maximum number of tokens processed from the input text.

Example

{
  "Semantic": {
    "Enabled": true,
    "DefaultProvider": "Internal",
    "Providers": {
      "Internal": {
        "ModelType": "Community",
        "PackagePath": "models/community-model.zip",
        "MaxLength": 128
      }
    }
  }
}

Pro Model Example

{
  "Semantic": {
    "Enabled": true,
    "DefaultProvider": "Internal",
    "Providers": {
      "Internal": {
        "ModelType": "Pro",
        "PackagePath": "models/pro-model.zip",
        "LicensePath": "/opt/synentra/licenses/license.json"
      }
    }
  }
}

---

OpenAI

Uses an OpenAI chat model to perform semantic classification.

Properties

Property	Type	Default	Description
ApiKey	string	""	OpenAI API key.
Model	string	"gpt-4o-mini"	OpenAI model identifier.
Temperature	double?	0.0	Sampling temperature. Lower values produce more deterministic classifications.
MaxTokens	int?	256	Maximum number of output tokens generated by the model.

Example

{
  "Semantic": {
    "Enabled": true,
    "DefaultProvider": "OpenAi",
    "Providers": {
      "OpenAi": {
        "ApiKey": "sk-...",
        "Model": "gpt-4o-mini",
        "Temperature": 0,
        "MaxTokens": 256
      }
    }
  }
}

---

Azure AI

Uses an Azure OpenAI deployment.

Properties

Property	Type	Default	Description
Endpoint	string	""	Azure OpenAI endpoint.
ApiKey	string	""	Azure API key.
Model	string	"gpt-4o-mini"	Azure deployment/model name.
Temperature	double?	0.0	Sampling temperature.
MaxTokens	int?	256	Maximum generated tokens.

Example

{
  "Semantic": {
    "Enabled": true,
    "DefaultProvider": "AzureAi",
    "Providers": {
      "AzureAi": {
        "Endpoint": "https://my-resource.openai.azure.com/",
        "ApiKey": "...",
        "Model": "gpt-4o-mini",
        "Temperature": 0,
        "MaxTokens": 256
      }
    }
  }
}

---

Google Gemini

Uses Google's Gemini models for semantic classification.

Properties

Property	Type	Default	Description
ApiKey	string	""	Gemini API key.
Model	string	"gemini-2.0-flash"	Gemini model identifier.
Temperature	double?	0.0	Sampling temperature.
MaxTokens	int?	256	Maximum generated tokens.

Example

{
  "Semantic": {
    "Enabled": true,
    "DefaultProvider": "Gemini",
    "Providers": {
      "Gemini": {
        "ApiKey": "...",
        "Model": "gemini-2.0-flash",
        "Temperature": 0,
        "MaxTokens": 256
      }
    }
  }
}

---

Ollama

Uses a locally hosted Ollama server.

Properties

Property	Type	Default	Description
Endpoint	string	"http://localhost:11434"	Ollama server endpoint.
Model	string	"llama3.2"	Local model name available in Ollama.

Example

{
  "Semantic": {
    "Enabled": true,
    "DefaultProvider": "Ollama",
    "Providers": {
      "Ollama": {
        "Endpoint": "http://localhost:11434",
        "Model": "llama3.2"
      }
    }
  }
}

---

Confidence Threshold

Every semantic provider returns:

• a classification
• a confidence score between 0 and 1

Synentra compares the returned confidence against ConfidenceThreshold.

Confidence	Result
≥ threshold	The semantic verdict is trusted.
< threshold + AllowLowConfidence = true	The request continues despite low confidence.
< threshold + AllowLowConfidence = false	The request is rejected or escalated according to the configured enforcement pipeline.

---

Evaluation Pipeline

When semantic analysis is enabled, Synentra performs the following steps:

1. Receives the proxied request.
2. Extracts relevant text from the request.
3. Sends the text to the configured semantic provider.
4. Receives the predicted intent and confidence score.
5. Compares the confidence against ConfidenceThreshold.
6. Combines the semantic result with deterministic policy evaluation.
7. Allows, denies, or escalates the request (for example, to Human-in-the-Loop) depending on the final policy decision.

---

Recommendations

Scenario	Recommended Provider
Offline / air-gapped deployments	Internal
Lowest latency	Internal
Maximum privacy	Internal
Highest reasoning capability	OpenAI or Azure AI
Google AI ecosystem	Gemini
Self-hosted LLMs	Ollama

---

Example Configuration

{
  "Semantic": {
    "Enabled": true,
    "ConfidenceThreshold": 0.8,
    "AllowLowConfidence": false,
    "DefaultProvider": "Internal",
    "Providers": {
      "Internal": {
        "ModelType": "Community",
        "PackagePath": "models/community-model.zip",
        "MaxLength": 128
      }
    }
  }
}