/**
 * @copyright Sister Software
 * @license AGPL-3.0
 * @author Teffen Ellis, et al.
 *
 *   Browser ONNX inference wrapper. Implements the same `NeuralRunner` contract `@mailwoman/neural`'s
 *   classifier consumes, but backed by `onnxruntime-web` (WASM + optional WebGPU) instead of
 *   `onnxruntime-node`.
 *
 *   Execution provider strategy:
 *
 *   - Try WebGPU first when `useWebGPU !== false`. ~10× faster than WASM on supported devices, but
 *       availability depends on browser (Chromium 113+, Safari Tech Preview) AND hardware. The
 *       runtime surfaces a clean error when WebGPU is unavailable, so the constructor falls back to
 *       WASM automatically.
 *   - WASM (SIMD when available) is the universal fallback. ~2× slower than WebGPU on the same model
 *       but works everywhere onnxruntime-web does — including in Node, which is how the test
 *       harness exercises this file.
 *
 *   Tensor shape + I/O contract matches `ONNXRunner` exactly: fixed-length int64 inputs, padded with
 *   zeros + attention_mask, output is a `logits` tensor of shape `[batch, seq, num_labels]`. See
 *   `@mailwoman/neural/onnx-runner` for the full export contract this file mirrors.
 */
import type { NeuralRunner } from "./classifier.ts";
import type { InferResult } from "./onnx-runner.ts";
export interface WebONNXRunnerOpts {
    /**
     * Try the WebGPU execution provider first. Defaults to true. Set false to skip the WebGPU probe — useful in test
     * environments where WebGPU isn't available and the probe failure adds latency.
     */
    useWebGPU?: boolean;
    /**
     * Fixed sequence length the model expects. Matches `ONNXRunner.DEFAULT_FIXED_SEQ_LEN` (128) by default. Re-quantized
     * models can override.
     */
    fixedSeqLen?: number;
    /**
     * Optional override for where onnxruntime-web should load its `.wasm` assets from. Defaults to the package's CDN
     * paths; bundlers usually want to point this at a self-hosted copy.
     *
     * Example: `setWASMPaths("/static/ort/")` and put the .wasm files at /static/ort/.
     */
    wasmPathsRoot?: string;
}
/**
 * Sequence length the web runtime pads to when the model was exported with a fixed input shape. WebGPU requires static
 * shapes, so a fixed length is the portable default.
 */
export declare const DEFAULT_FIXED_SEQ_LEN = 128;
export interface WebONNXRunnerDiagnostics {
    backend: "webgpu" | "wasm";
    modelBytes: number;
}
export declare class WebONNXRunner implements NeuralRunner {
    #private;
    readonly fixedSeqLen: number;
    diagnostics: WebONNXRunnerDiagnostics | null;
    private readonly modelBytes;
    private readonly opts;
    private constructor();
    /**
     * Construct from already-fetched model bytes.
     */
    static fromBytes(modelBytes: Uint8Array, opts?: WebONNXRunnerOpts): Promise<WebONNXRunner>;
    /**
     * Fetch the model from a URL and construct.
     */
    static fromURL(modelURL: string, opts?: WebONNXRunnerOpts): Promise<WebONNXRunner>;
    /**
     * Names of the inputs the loaded ONNX graph declares. `null` until the session has been created (first `infer()`
     * call). Lets callers (e.g. the neural-web loader) detect anchor/gazetteer-trained models and warn loudly when the
     * corresponding feature source wasn't provided — running such a model on the zero-filled fallback is the measured
     * train/inference mismatch ("the zero-fill trap"), not a quality-neutral degrade.
     */
    get inputNames(): readonly string[] | null;
    infer(tokenIDs: number[], anchor?: {
        features: ReadonlyArray<ReadonlyArray<number>>;
        confidence: ReadonlyArray<number>;
    }, gazetteer?: {
        features: ReadonlyArray<ReadonlyArray<number>>;
        confidence: ReadonlyArray<number>;
    }, country?: {
        features: ReadonlyArray<ReadonlyArray<number>>;
        confidence: ReadonlyArray<number>;
    }, evidence?: {
        streetType?: {
            features: ReadonlyArray<ReadonlyArray<number>>;
            confidence: ReadonlyArray<number>;
        };
        localitySurface?: {
            features: ReadonlyArray<ReadonlyArray<number>>;
            confidence: ReadonlyArray<number>;
        };
    }): Promise<InferResult>;
}
//# sourceMappingURL=web-onnx-runner.d.ts.map