Branch on relevance

Decide whether a passage answers a search query, then use an ordinary Zig if to include or omit it. Use Decisions when you need a fixed boolean answer rather than extracted text.

Apply a probability threshold

Pass a configured common Client for OpenAI with gpt-6-luna, or a TypeSafe-compatible service. See Decisions provider configuration.

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

fn checkRelevance(client: *instructor.Client) !void {
    const input =
        \\Search query: How do I reset my password?
        \\Passage: Select 'Forgot password' to receive a reset link.
    ;
    const relevant = try client.decisions.condition(.{
        .input = input,
        .instructions = "Does the passage answer the search query?",
        .threshold = 0.8,
    });

    if (relevant) {
        std.debug.print("Include the passage.\n", .{});
    } else {
        std.debug.print("Leave the passage out.\n", .{});
    }
}

Choose the threshold from examples

The result is true when the reported predicate probability is at least 0.8. The probability is a model estimate, not a guarantee that the passage is relevant. Choose a threshold using labeled examples from your application. Without an explicit threshold, the default is 0.5.

Keep both the query and passage in the input. Keep the fixed question in instructions, separate from the text being judged.

condition frees the response internally and returns a plain bool. There is no result to deinitialize. The caller owns the client. Errors propagate through try; a failed request is not silently treated as false.

Full source