AgentFrameworkToolkit.Jev 1.24.2

AgentFrameworkToolkit.Jev

Evaluate text with the Jev Decisions API using the same requests, attributes, results, and IDecisionFactory contract as OpenAI. This package contains its own HTTP implementation, adapted from JevDotNet, and has no dependency on the JevDotNet package.

Usage

Supply an API key and model explicitly. There is no default model.

using AgentFrameworkToolkit.Decisions;
using AgentFrameworkToolkit.Jev;

IDecisionFactory factory = new JevDecisionFactory(apiKey, decisionModel);

bool damaged = await factory.IsTrueAsync(new ProbabilityRequest
{
    Input = "The screen arrived shattered.",
    Question = "Is the product damaged?",
    Threshold = 0.7
});

DecisionResponse<SupportDecision> response =
    await factory.CreateDecisionAsync<SupportDecision>(new DecisionRequest
    {
        Input = "The screen arrived shattered."
    });

public class SupportDecision
{
    [ProbabilityQuestion("Is the product damaged?", threshold: 0.7)]
    public bool IsDamaged { get; set; }

    [ChoiceQuestion<Department>("Which department should handle this?")]
    public Department Department { get; set; }

    [ScoreQuestion<Severity>("How severe is the damage?")]
    public Score<Severity>? Severity { get; set; }
}

public enum Department { Support, Returns }
public enum Severity { Minor, Moderate, Severe }

ChooseAsync<TEnum> returns an enum choice. ProbabilityAsync returns a Probability whose IsTrue uses the request threshold. ScoreAsync<TEnum> returns the weighted score, confidence, probabilities, and level descriptions. For multiple questions or token usage, use CreateDecisionAsync<T>.

For runtime-defined questions, use the non-generic CreateDecisionAsync overload with DynamicDecisionRequest. Supply a list of ChoiceQuestion, ScoreQuestion, and ProbabilityQuestion objects, then retrieve answers directly through GetChoice(id), GetScore(id), and GetProbability(id) on the response. IDs and lookup are case-insensitive; unknown IDs and mismatched accessors throw. See the shared usage example. The existing text-only restriction and limit of 2–10 score levels also apply to dynamic requests.

Predicate properties support bool, double, decimal, their nullable forms, and Probability. Numeric probabilities ignore the attribute threshold. Choice properties support the enum, its nullable form, or Choice<TEnum>. Score properties support double, decimal, their nullable forms, or Score<TEnum>. Enum numeric order defines zero-based levels, and DescriptionAttribute supplies criteria. Jev score questions require 2–10 levels. Question properties require public setters. Refused, missing, or invalid answers throw; no partial result is returned.

Responses include the returned model and input, output, and total token counts. Total tokens are calculated from Jev's input and output counts.

Connection and dependency injection

JevConnection connection = new(apiKey)
{
    Endpoint = new Uri("https://api.typesafe.ai/v1/systemone"),
    NetworkTimeout = TimeSpan.FromSeconds(60),
    HttpClientFactory = () => httpClient
};

JevDecisionFactory factory = new(connection, decisionModel);
services.AddJevDecisionFactory(connection, decisionModel);
// Or: services.AddJevDecisionFactory(apiKey, decisionModel);

The factory retains its connection settings and HTTP client at construction. The default endpoint is Jev's /v1/systemone endpoint. A supplied HTTP client remains caller-owned. NetworkTimeout applies per request without changing the shared HTTP client's timeout. Requests accept cancellation tokens. Set RawHttpCallDetails on a request to inspect its URL and request/response bodies, including HTTP error responses.

DI registers JevDecisionFactory and IDecisionFactory as the same singleton. When registering multiple decision providers, resolve their concrete factories; the last registration supplies IDecisionFactory.

Current limitations

All image overloads are present for contract compatibility and throw NotSupportedException until Jev supports images. SafetyIdentifier is unsupported by Jev and throws when supplied.

Migration from OpenAI-specific contracts

Use AgentFrameworkToolkit.Decisions instead of AgentFrameworkToolkit.OpenAI.Decisions, and DecisionResponse<T> instead of OpenAIDecisionResponse<T>. Your attributed result classes and requests can then be used with either decision provider.