TruthVouch.SemanticKernel 1.0.1

TruthVouch.SemanticKernel

Semantic Kernel integration for TruthVouch — verify LLM outputs for factual accuracy and attach trust scores to kernel function results.

Installation

dotnet add package TruthVouch.SemanticKernel

Quick Start

Register TruthVouch directly on your IKernelBuilder. This wires in both the function invocation filter and the prompt render filter automatically:

using TruthVouch.SemanticKernel;

var kernel = Kernel.CreateBuilder()
    .AddOpenAIChatCompletion("gpt-4o", openAiApiKey)
    .AddTruthVouch(o =>
    {
        o.ApiKey         = "your-truthvouch-api-key";
        o.TrustThreshold = 0.75;
        o.ThrowOnLowTrust = true;
    })
    .Build();

If you use a plain IServiceCollection (e.g., ASP.NET Core), call the IServiceCollection overload instead:

builder.Services.AddTruthVouch(o =>
{
    o.ApiKey         = "your-truthvouch-api-key";
    o.TrustThreshold = 0.8;
});

Components

TruthVouchPlugin

A Semantic Kernel plugin that exposes fact-verification as a first-class kernel function (verify_content). The function accepts a text string and an optional verification mode, calls the TruthVouch Trust API, and returns a structured VerifyResponse containing the trust score and per-claim verdicts.

Add the plugin to your kernel to let the AI planner invoke it on demand:

kernel.Plugins.AddFromObject(kernel.Services.GetRequiredService<TruthVouchPlugin>(), "TruthVouch");

TruthVouchFilter

Implements IFunctionInvocationFilter. After every kernel function call that returns a string, this filter automatically sends the output to the Trust API and:

• Attaches truthvouch_trust_score and truthvouch_verification_id to the FunctionResult metadata on every call.
• When the score falls below TrustThreshold and ThrowOnLowTrust is false, logs a warning and adds a truthvouch_warning key to the result metadata.
• When ThrowOnLowTrust is true, throws a TrustVerificationException.

Functions decorated with [SkipTruthVerification] are bypassed entirely.

TruthVouchPromptFilter

Implements IPromptRenderFilter. When VerifyInputPrompts is true, this filter verifies the fully rendered prompt text before it is sent to the LLM. The trust score and verification ID are attached to the prompt context arguments (truthvouch_prompt_trust_score, truthvouch_prompt_verification_id) so downstream filters can inspect them without a second API call.

Both filters are registered automatically when you call AddTruthVouch on an IKernelBuilder.

Configuration

All options are set via the TruthVouchOptions delegate passed to AddTruthVouch.

License

Apache-2.0