// Speech models offered in the picker, and the per-device weights they need.
//
// The two backends want different quantizations, so they download different
// files and the picker has to quote the size for whichever one will actually
// run:
//
//   wasm   — int8 encoder + int8 merged decoder. Small, CPU-friendly.
//   webgpu — fp32 encoder + 4-bit merged decoder. This is the combination the
//            transformers.js WebGPU examples use; int8 matmul is poorly served
//            on WebGPU. It runs far faster but downloads 2–3× as much, because
//            the 4-bit decoder only quantizes matmul weights and leaves the
//            embeddings at full precision.
//
// `mb` figures are the encoder + decoder + tokenizer/config bytes, taken from
// the HuggingFace file listing. Everything is cached after the first visit.
export const MODELS = [
  {
    id: 'onnx-community/moonshine-tiny-ONNX',
    label: 'Moonshine Tiny',
    mb: { wasm: 32, webgpu: 79 },
    note: 'fastest; lighter punctuation',
  },
  {
    id: 'onnx-community/whisper-tiny.en',
    label: 'Whisper Tiny',
    mb: { wasm: 44, webgpu: 122 },
    note: 'quick; less accurate on hard words',
  },
  {
    id: 'onnx-community/moonshine-base-ONNX',
    label: 'Moonshine Base',
    mb: { wasm: 67, webgpu: 157 },
    note: 'fast and accurate; lighter punctuation',
  },
  {
    id: 'onnx-community/whisper-base.en',
    label: 'Whisper Base',
    mb: { wasm: 80, webgpu: 209 },
    note: 'good accuracy and punctuation',
  },
  {
    id: 'onnx-community/whisper-small.en',
    label: 'Whisper Small',
    mb: { wasm: 252, webgpu: 588 },
    note: 'most accurate; slow without WebGPU',
  },
];

export const DEFAULT_MODEL = 'onnx-community/whisper-base.en';

// Quantization suffixes per device — these decide both which files
// transformers.js fetches and which ones we look for in the cache.
export const SUFFIX = {
  wasm: { encoder: '_quantized', decoder: '_quantized' },
  webgpu: { encoder: '', decoder: '_q4' },
};

export async function detectDevice() {
  if (!navigator.gpu) return 'wasm';
  try {
    return (await navigator.gpu.requestAdapter()) ? 'webgpu' : 'wasm';
  } catch {
    return 'wasm';
  }
}

// The two large files for a model on a given device. Their presence in the
// cache is what "downloaded" in the picker means.
const weightFiles = (id, device) => {
  const s = SUFFIX[device];
  return [
    `https://huggingface.co/${id}/resolve/main/onnx/encoder_model${s.encoder}.onnx`,
    `https://huggingface.co/${id}/resolve/main/onnx/decoder_model_merged${s.decoder}.onnx`,
  ];
};

export async function cachedModels(device) {
  const cached = new Set();
  if (!self.caches) return cached;
  try {
    const cache = await caches.open('transformers-cache');
    await Promise.all(
      MODELS.map(async (m) => {
        const hits = await Promise.all(weightFiles(m.id, device).map((u) => cache.match(u)));
        if (hits.every(Boolean)) cached.add(m.id);
      }),
    );
  } catch { /* cache unavailable (private mode, etc.) — report none cached */ }
  return cached;
}