Pervaxis.Core.AI 1.6.0

Pervaxis.Core.AI

AWS Bedrock Converse API integration with guardrail evaluation, human approval gating, versioned prompt registry, and fully traceable AI invocation results. Implements Section 18 of the Pervaxis Platform Spec.

Installation

dotnet add package Pervaxis.Core.AI

Registration

builder.Services.AddPervaxisAI(options =>
{
    options.Bedrock.DefaultModelId = BedrockModelIds.ClaudeSonnet46;
    options.Bedrock.Region = "us-east-1";
});

Configuration via appsettings.json:

{
  "Pervaxis": {
    "AI": {
      "Bedrock": {
        "DefaultModelId": "us.anthropic.claude-sonnet-4-6-20251101-v1:0",
        "Region": "us-east-1",
        "MaxTokens": 2048,
        "Temperature": 0.3
      }
    }
  }
}

Available Models

BedrockModelIds.ClaudeSonnet46   // us.anthropic.claude-sonnet-4-6-20251101-v1:0
BedrockModelIds.ClaudeOpus46     // us.anthropic.claude-opus-4-6-20251101-v1:0
BedrockModelIds.ClaudeHaiku45    // us.anthropic.claude-haiku-4-5-20251001-v1:0
BedrockModelIds.NovaProV1        // us.amazon.nova-pro-v1:0
BedrockModelIds.TitanEmbeddingsV2 // amazon.titan-embed-text-v2:0

Invoking Bedrock

public class TriageService(IBedrockClient bedrock, IPromptRegistry prompts)
{
    public async Task<string> AnalyseAsync(string incidentDetail, CancellationToken ct)
    {
        var prompt = prompts.GetLatest("triage.root-cause-analysis");

        var result = await bedrock.InvokeAsync(new AiInvocationRequest
        {
            SystemPrompt = prompt.Render(new Dictionary<string, string>
            {
                ["serviceName"] = "order-service"
            }),
            UserMessage = incidentDetail,
            PromptKey = prompt.Key,
            PromptVersion = prompt.Version,
            RiskLevel = AiRiskLevel.Low,
            CorrelationId = correlationId
        }, ct);

        return result.Output;
    }
}

AI Invocation Pipeline (Section 18)

Every invocation flows through:

Request → GuardrailPolicy.EvaluateAsync()
             ↓ Blocked → throw
        → ApprovalGate.RequiresApproval(riskLevel)?
             ↓ Yes → ApprovalGate.RequestApprovalAsync()
        → BedrockRuntime.ConverseAsync()
        → AiInvocationResult<string> (fully traceable)

Versioned Prompt Registry

// Register prompts at startup
registry.Register(new PromptDefinition
{
    Key = "triage.root-cause-analysis",
    Version = "1.0.0",
    Content = "You are a {serviceName} expert. Analyse: {incident}",
    RiskLevel = AiRiskLevel.Low
});

// Retrieve by exact version or latest
var prompt = registry.Get("triage.root-cause-analysis", "1.0.0");
var latest = registry.GetLatest("triage.root-cause-analysis");

// Render with variables
var rendered = prompt.Render(new Dictionary<string, string>
{
    ["serviceName"] = "order-service",
    ["incident"] = "DB connection timeout"
});

Risk Levels and Approval Gate

For production, register a real IAiActionApprovalGate (ServiceNow, Slack workflow, etc.):

builder.Services.AddScoped<IAiActionApprovalGate, ServiceNowApprovalGate>();

Custom Guardrails

public sealed class ContentFilterPolicy : IAiGuardrailPolicy
{
    public string PolicyName => "content-filter";

    public Task<GuardrailResult> EvaluateAsync(
        AiInvocationRequest request, CancellationToken ct = default)
    {
        if (request.UserMessage.Contains("DROP TABLE"))
            return Task.FromResult(GuardrailResult.Block(PolicyName, "SQL injection attempt"));

        return Task.FromResult(GuardrailResult.Pass());
    }
}

// Register
builder.Services.AddAiGuardrailPolicy<ContentFilterPolicy>();

AWS IAM Requirements

{
  "Effect": "Allow",
  "Action": ["bedrock:InvokeModel", "bedrock:Converse"],
  "Resource": "arn:aws:bedrock:us-east-1::foundation-model/*"
}

Pervaxis Platform · Clarivex Technologies · https://clarivex.tech