tai

What it does

TypeSafe models such as Jev answer questions about state with calibrated probabilities instead of generated text. There are three kinds of question:

• noul: a yes/no question. The answer is the probability of yes.
• choice: pick one option from a set you define.
• score: rate against ordered levels you define.

tai wraps the HTTP API with a typed DSL. You describe your questions as a Zig struct literal, and tai works out the result type at compile time. Choice options become a real enum, so a switch on the answer is checked for exhaustiveness.

const r = try client.ask("Help! My payouts have been failing for 3 days.", .{
    .is_urgent = tai.noul("Does this convey urgency?"),
    .department = tai.choice("Which team should handle this?", .{
        .billing = "Payments, invoicing, refunds",
        .technical = "Bugs, outages, integrations",
        .sales = "Pricing, upgrades, new accounts",
    }),
    .frustration = tai.score("How frustrated is the customer?", .{ "Calm", "Frustrated", "Very angry" }),
}, .{});

const page_oncall = r.answers.is_urgent > 0.8;
const queue = switch (r.answers.department.choice) {
    .billing => "payments",
    .technical => "engineering",
    .sales => "sales",
};

Install

tai targets Zig 0.17.0-dev and has no dependencies beyond the standard library.

zig fetch --save git+https://github.com/benomahony/zig-typesafe-ai

Then, in build.zig:

const tai = b.dependency("tai", .{ .target = target, .optimize = optimize });
exe.root_module.addImport("tai", tai.module("tai"));

Quick start

Get an API key from TypeSafe and export it:

export TYPESAFE_API_KEY=...

const std = @import("std");
const tai = @import("tai");

pub fn main(init: std.process.Init) !void {
    var client: tai.Client = try .init(init.gpa, init.io, .{ .environ_map = init.environ_map });
    defer client.deinit();

    const r = try client.ask("I was charged twice. Please help.", .{
        .billing = tai.noul("Is this about billing?"),
    }, .{});

    std.debug.print("billing: {d:.2} ({s})\n", .{ r.answers.billing, r.model() });
}

ask returns a plain value that owns no heap memory, so there is nothing to free.

Every example is a test

Every Zig example on this site is quoted from an end-to-end test that runs against the live API (tests/e2e.zig), or from a real downstream project (tests/consumer) that depends on tai exactly as shown above. If an example stops working, the build fails.

Where next

• Typed questions covers the full DSL: criteria, enums, structured state and the result types.
• Client & errors covers configuration, retries, error handling and listing models.
• Runtime questions is for when the set of questions isn’t known until runtime.