OmniRoute/open-sse/handlers/audioTranscription.ts

import { CORS_HEADERS } from "../utils/cors.ts";
import { Buffer } from "node:buffer";
/**
 * Audio Transcription Handler
 *
 * Handles POST /v1/audio/transcriptions (Whisper API format).
 * Proxies multipart/form-data to upstream providers.
 *
 * Supported provider formats:
 * - OpenAI/Groq/Qwen3: standard multipart form-data proxy
 * - Deepgram: raw binary audio POST with model via query param
 * - AssemblyAI: async workflow (upload → submit → poll)
 * - Gladia: async workflow (upload → submit pre-recorded job → poll result_url)
 * - Nvidia NIM: multipart POST, transform response to { text }
 * - HuggingFace Inference: POST raw binary to /models/{model_id}
 */

import {
  getTranscriptionProvider,
  parseTranscriptionModel,
  type AudioProvider,
} from "../config/audioRegistry.ts";
import { buildAuthHeaders } from "../config/registryUtils.ts";
import { kieExecutor } from "../executors/kie.ts";
import { vertexTranscribe } from "../executors/vertexMedia.ts";
import { errorResponse } from "../utils/error.ts";
import { handleOpenRouterTranscription } from "./openrouterTranscription.ts";

type TranscriptionCredentials = {
  apiKey?: string;
  accessToken?: string;
};

/**
 * Return a CORS error response from an upstream fetch failure
 */
export function upstreamErrorResponse(res, errText) {
  // Always return JSON so the client can parse the error reliably
  let errorMessage: string;
  try {
    const parsed = JSON.parse(errText);
    // Guard against `parsed.error` or `parsed.detail` being objects
    const raw =
      parsed?.err_msg ||
      parsed?.error?.message ||
      (typeof parsed?.error === "string" ? parsed.error : null) ||
      parsed?.message ||
      (typeof parsed?.detail === "string" ? parsed.detail : parsed?.detail?.message) ||
      null;
    errorMessage = raw ? String(raw) : errText || `Upstream error (${res.status})`;
  } catch {
    errorMessage = errText || `Upstream error (${res.status})`;
  }

  return Response.json(
    { error: { message: errorMessage, code: res.status } },
    {
      status: res.status,
      headers: { ...CORS_HEADERS },
    }
  );
}

/**
 * Validate a path segment to prevent path traversal / SSRF.
 */
function isValidPathSegment(segment: string): boolean {
  return !segment.includes("..") && !segment.includes("//");
}

function getUploadedFileName(file: Blob & { name?: unknown }): string {
  return typeof file.name === "string" && file.name.length > 0 ? file.name : "audio.wav";
}

/**
 * `body` is `Uint8Array<ArrayBuffer>`, not bare `Uint8Array`: `new Uint8Array(n)`
 * is always ArrayBuffer-backed, and only that narrower form satisfies `BodyInit`
 * (the bare type widens to `ArrayBufferLike`, which admits `SharedArrayBuffer`).
 */
export async function buildMultipartBody(
  file: Blob & { name?: unknown },
  fields: Record<string, string>,
  fileFieldName = "file"
): Promise<{ body: Uint8Array<ArrayBuffer>; contentType: string }> {
  const boundary = "----OmniRouteAudioBoundary" + Date.now().toString(36);
  const parts: Uint8Array[] = [];
  const encoder = new TextEncoder();

  for (const [name, value] of Object.entries(fields)) {
    parts.push(
      encoder.encode(
        `--${boundary}\r\nContent-Disposition: form-data; name="${name}"\r\n\r\n${value}\r\n`
      )
    );
  }

  const fileName = getUploadedFileName(file)
    .replace(/["]/g, "_")
    .replace(/[\r\n]/g, "_");
  const fileBytes = new Uint8Array(await file.arrayBuffer());
  parts.push(
    encoder.encode(
      `--${boundary}\r\nContent-Disposition: form-data; name="${fileFieldName}"; filename="${fileName}"\r\nContent-Type: ${file.type || "application/octet-stream"}\r\n\r\n`
    )
  );
  parts.push(fileBytes);
  parts.push(encoder.encode(`\r\n--${boundary}--\r\n`));

  const totalLength = parts.reduce((sum, p) => sum + p.length, 0);
  const body = new Uint8Array(totalLength);
  let offset = 0;
  for (const part of parts) {
    body.set(part, offset);
    offset += part.length;
  }

  return { body, contentType: "multipart/form-data; boundary=" + boundary };
}

/**
 * Infer a suitable Content-Type for Deepgram from the browser-provided MIME
 * type and the original filename.  Deepgram accepts `audio/*` and many raw
 * formats, but `video/*` causes it to silently fail with "no speech detected".
 *
 * Strategy:
 * 1. If the browser says `audio/*`, keep it as-is.
 * 2. If it's `video/*` (e.g. `.mp4`), remap to the audio equivalent so
 *    Deepgram extracts the audio track.  `.mp4` → `audio/mp4`, etc.
 * 3. Fall back to `application/octet-stream` which tells Deepgram to
 *    auto-detect from the raw bytes (most reliable for unknown formats).
 */
function resolveAudioContentType(file: Blob & { name?: unknown }): string {
  const browserType = (file.type || "").toLowerCase();
  const fileName = typeof file.name === "string" ? file.name.toLowerCase() : "";

  // 1) Browser already says it's audio — trust it
  if (browserType.startsWith("audio/")) return browserType;

  // 2) Derive from file extension (covers video/* and empty MIME)
  const ext = fileName.includes(".") ? fileName.split(".").pop() : "";
  const EXT_TO_MIME: Record<string, string> = {
    mp3: "audio/mpeg",
    mp4: "audio/mp4",
    m4a: "audio/mp4",
    wav: "audio/wav",
    ogg: "audio/ogg",
    flac: "audio/flac",
    webm: "audio/webm",
    aac: "audio/aac",
    wma: "audio/x-ms-wma",
    opus: "audio/opus",
  };
  if (ext && EXT_TO_MIME[ext]) return EXT_TO_MIME[ext];

  // 3) Fallback — let Deepgram auto-detect from raw bytes
  return "application/octet-stream";
}

/**
 * Handle Deepgram transcription (raw binary audio, model via query param)
 */
async function handleDeepgramTranscription(
  providerConfig,
  file,
  modelId,
  token,
  formData?: FormData
) {
  const url = new URL(providerConfig.baseUrl);
  url.searchParams.set("model", modelId);
  url.searchParams.set("smart_format", "true");
  url.searchParams.set("punctuate", "true");

  // Language: if caller specified one, use it; otherwise let Deepgram auto-detect
  const langParam = formData?.get("language");
  if (typeof langParam === "string" && langParam.trim()) {
    url.searchParams.set("language", langParam.trim());
  } else {
    url.searchParams.set("detect_language", "true");
  }

  const arrayBuffer = await file.arrayBuffer();

  const res = await fetch(url.toString(), {
    method: "POST",
    headers: {
      ...buildAuthHeaders(providerConfig, token),
      "Content-Type": resolveAudioContentType(file),
    },
    body: arrayBuffer,
  });

  if (!res.ok) {
    return upstreamErrorResponse(res, await res.text());
  }

  const data = await res.json();
  // Transform Deepgram response to OpenAI Whisper format
  const text = data.results?.channels?.[0]?.alternatives?.[0]?.transcript ?? null;

  // null means the audio had no recognizable speech (music, silence, etc.)
  // Return it explicitly so the client can distinguish from a credentials error
  return Response.json(
    { text: text ?? "", noSpeechDetected: text === null || text === "" },
    { headers: { ...CORS_HEADERS } }
  );
}

/**
 * Handle AssemblyAI transcription (async: upload file → submit → poll)
 */
async function handleAssemblyAITranscription(providerConfig, file, modelId, token) {
  const authHeaders = buildAuthHeaders(providerConfig, token);

  // Step 1: Upload the audio file
  const arrayBuffer = await file.arrayBuffer();
  const uploadRes = await fetch("https://api.assemblyai.com/v2/upload", {
    method: "POST",
    headers: {
      ...authHeaders,
      "Content-Type": "application/octet-stream",
    },
    body: arrayBuffer,
  });

  if (!uploadRes.ok) {
    return upstreamErrorResponse(uploadRes, await uploadRes.text());
  }

  const { upload_url } = await uploadRes.json();

  // Step 2: Submit transcription request
  const submitRes = await fetch(providerConfig.baseUrl, {
    method: "POST",
    headers: {
      ...authHeaders,
      "Content-Type": "application/json",
    },
    body: JSON.stringify({
      audio_url: upload_url,
      speech_models: [modelId],
      language_detection: true,
    }),
  });

  if (!submitRes.ok) {
    return upstreamErrorResponse(submitRes, await submitRes.text());
  }

  const { id: transcriptId } = await submitRes.json();

  // Step 3: Poll for completion (max 120s)
  const pollUrl = `${providerConfig.baseUrl}/${transcriptId}`;
  const maxWait = 120_000;
  const start = Date.now();

  while (Date.now() - start < maxWait) {
    await new Promise((r) => setTimeout(r, 2000));

    const pollRes = await fetch(pollUrl, { headers: authHeaders });
    if (!pollRes.ok) continue;

    const result = await pollRes.json();

    if (result.status === "completed") {
      return Response.json({ text: result.text || "" }, { headers: { ...CORS_HEADERS } });
    }

    if (result.status === "error") {
      return errorResponse(500, result.error || "AssemblyAI transcription failed");
    }
  }

  return errorResponse(504, "AssemblyAI transcription timed out after 120s");
}

/**
 * Handle Gladia transcription (async: upload file → submit pre-recorded job → poll result_url)
 */
async function handleGladiaTranscription(providerConfig, file, modelId, token) {
  const authHeaders = buildAuthHeaders(providerConfig, token);

  // Step 1: Upload the audio file (multipart/form-data)
  const { body: uploadBody, contentType: uploadCT } = await buildMultipartBody(file, {});
  const uploadRes = await fetch("https://api.gladia.io/v2/upload", {
    method: "POST",
    headers: { ...authHeaders, "Content-Type": uploadCT },
    body: uploadBody,
  });

  if (!uploadRes.ok) {
    return upstreamErrorResponse(uploadRes, await uploadRes.text());
  }

  const { audio_url } = await uploadRes.json();

  // Step 2: Submit the pre-recorded transcription job
  const submitRes = await fetch(providerConfig.baseUrl, {
    method: "POST",
    headers: { ...authHeaders, "Content-Type": "application/json" },
    body: JSON.stringify({ audio_url, model: modelId }),
  });

  if (!submitRes.ok) {
    return upstreamErrorResponse(submitRes, await submitRes.text());
  }

  const { result_url: resultUrl } = await submitRes.json();
  if (!resultUrl) {
    return errorResponse(502, "Gladia did not return a result_url");
  }

  // Step 3: Poll for completion (max 120s)
  const maxWait = 120_000;
  const start = Date.now();

  while (Date.now() - start < maxWait) {
    await new Promise((r) => setTimeout(r, 2000));

    const pollRes = await fetch(resultUrl, { headers: authHeaders });
    if (!pollRes.ok) continue;

    const result = await pollRes.json();

    if (result.status === "done") {
      const text = result.result?.transcription?.full_transcript || "";
      return Response.json({ text }, { headers: { ...CORS_HEADERS } });
    }

    if (result.status === "error") {
      return errorResponse(500, result.error_code || result.error || "Gladia transcription failed");
    }
  }

  return errorResponse(504, "Gladia transcription timed out after 120s");
}

/**
 * Handle Nvidia NIM transcription
 * Multipart POST, transform response to { text }
 */
async function handleNvidiaTranscription(providerConfig, file, modelId, token) {
  const { body, contentType } = await buildMultipartBody(file, { model: modelId });

  const res = await fetch(providerConfig.baseUrl, {
    method: "POST",
    headers: { ...buildAuthHeaders(providerConfig, token), "Content-Type": contentType },
    body,
  });

  if (!res.ok) {
    return upstreamErrorResponse(res, await res.text());
  }

  const data = await res.json();
  // Normalize to { text } — Nvidia may return { text } directly or nested
  const text = data.text || data.transcript || "";

  return Response.json({ text }, { headers: { ...CORS_HEADERS } });
}

/**
 * Handle HuggingFace Inference transcription
 * POST raw binary audio to {baseUrl}/{model_id}, returns { text }
 */
async function handleHuggingFaceTranscription(providerConfig, file, modelId, token) {
  if (!isValidPathSegment(modelId)) {
    return errorResponse(400, "Invalid model ID");
  }
  const url = `${providerConfig.baseUrl}/${modelId}`;
  const arrayBuffer = await file.arrayBuffer();

  const res = await fetch(url, {
    method: "POST",
    headers: {
      ...buildAuthHeaders(providerConfig, token),
      "Content-Type": resolveAudioContentType(file),
    },
    body: arrayBuffer,
  });

  if (!res.ok) {
    return upstreamErrorResponse(res, await res.text());
  }

  const data = await res.json();
  // HuggingFace returns { text } directly
  const text = data.text || "";

  return Response.json({ text }, { headers: { ...CORS_HEADERS } });
}

/**
 * Handle Kie.ai transcription
 */
async function handleKieAudioTranscription(providerConfig, file, modelId, token) {
  const baseUrl = providerConfig.baseUrl.replace(/\/$/, "");
  const fileBuffer = await file.arrayBuffer();
  const fileBase64 = Buffer.from(fileBuffer).toString("base64");
  let data;
  try {
    data = await kieExecutor.createTask({
      baseUrl,
      token,
      payload: {
        model: modelId,
        input: {
          file_name: getUploadedFileName(file),
          file_base64: fileBase64,
        },
      },
    });
  } catch (err: unknown) {
    const status =
      typeof err === "object" && err !== null && "status" in err
        ? Number((err as { status?: unknown }).status) || 502
        : 502;
    return Response.json(
      {
        error: {
          message: err instanceof Error ? err.message : "Kie transcription createTask failed",
          code: status,
        },
      },
      {
        status,
        headers: { ...CORS_HEADERS },
      }
    );
  }
  const taskId = data?.data?.taskId || data?.taskId;

  if (taskId) {
    return pollKieTranscriptionResult(baseUrl, modelId, taskId, token);
  }

  return Response.json(
    { text: data?.data?.text || data?.text || "" },
    { headers: { ...CORS_HEADERS } }
  );
}

/**
 * Internal polling for Kie.ai async transcription tasks
 */
async function pollKieTranscriptionResult(baseUrl, modelId, taskId, token) {
  void modelId;
  const statusUrl = kieExecutor.getTaskStatusUrl(baseUrl);
  try {
    const { data, state } = await kieExecutor.pollTask({
      statusUrl,
      taskId: String(taskId),
      token,
      timeoutMs: 120000,
      pollIntervalMs: 2000,
    });

    if (state === "success") {
      const text =
        data?.data?.response?.text ||
        data?.data?.resultText ||
        data?.data?.text ||
        data?.text ||
        "";
      return Response.json({ text }, { headers: { ...CORS_HEADERS } });
    }
  } catch (err: unknown) {
    const status =
      typeof err === "object" && err !== null && "status" in err
        ? Number((err as { status?: unknown }).status) || 504
        : 504;
    return errorResponse(
      status,
      err instanceof Error ? err.message : "Kie transcription generation timed out or failed"
    );
  }

  return errorResponse(504, "Kie transcription generation timed out or failed");
}

/**
 * Handle Rev AI transcription (async: submit job with media upload → poll → fetch transcript)
 *
 * Rev AI accepts the audio file directly in the job-submission multipart body
 * (field "media"), avoiding AssemblyAI's separate upload step. Once the job
 * reaches a terminal state we fetch the plain-text transcript.
 */
async function handleRevAiTranscription(providerConfig, file, modelId, token) {
  const authHeaders = buildAuthHeaders(providerConfig, token);
  const baseUrl = providerConfig.baseUrl.replace(/\/$/, "");

  // Step 1: submit the job — multipart body with "media" (file) + "options" (JSON)
  const options = JSON.stringify({ transcriber: modelId });
  const { body, contentType } = await buildMultipartBody(file, { options }, "media");

  const submitRes = await fetch(`${baseUrl}/jobs`, {
    method: "POST",
    headers: { ...authHeaders, "Content-Type": contentType },
    body,
  });

  if (!submitRes.ok) {
    return upstreamErrorResponse(submitRes, await submitRes.text());
  }

  const { id: jobId } = await submitRes.json();

  // Step 2: poll for completion (max 120s)
  const jobUrl = `${baseUrl}/jobs/${jobId}`;
  const maxWait = 120_000;
  const start = Date.now();

  while (Date.now() - start < maxWait) {
    await new Promise((r) => setTimeout(r, 2000));

    const pollRes = await fetch(jobUrl, { headers: authHeaders });
    if (!pollRes.ok) continue;

    const result = await pollRes.json();

    if (result.status === "transcribed") {
      const transcriptRes = await fetch(`${jobUrl}/transcript`, {
        headers: { ...authHeaders, Accept: "text/plain" },
      });
      if (!transcriptRes.ok) {
        return upstreamErrorResponse(transcriptRes, await transcriptRes.text());
      }
      const text = await transcriptRes.text();
      return Response.json({ text: text || "" }, { headers: { ...CORS_HEADERS } });
    }

    if (result.status === "failed") {
      return errorResponse(500, result.failure_detail || "Rev AI transcription failed");
    }
  }

  return errorResponse(504, "Rev AI transcription timed out after 120s");
}

/**
 * Speechmatics operating point (accuracy tier). Catalog model ids are the
 * real Speechmatics `operating_point` values ("standard", "enhanced",
 * "melia-1"), so this passes straight through — kept as a named seam in
 * case a future catalog id needs remapping.
 */
function speechmaticsOperatingPoint(modelId: string): string {
  return modelId;
}

/**
 * Fetch and return the finished Speechmatics transcript once a job reaches
 * the "done" state.
 */
async function fetchSpeechmaticsTranscript(jobUrl, authHeaders) {
  const transcriptRes = await fetch(`${jobUrl}/transcript?format=txt`, {
    headers: { ...authHeaders, Accept: "text/plain" },
  });
  if (!transcriptRes.ok) {
    return upstreamErrorResponse(transcriptRes, await transcriptRes.text());
  }
  const text = await transcriptRes.text();
  return Response.json({ text: text || "" }, { headers: { ...CORS_HEADERS } });
}

function speechmaticsJobErrorMessage(result): string {
  const errors = result?.job?.errors;
  const first = Array.isArray(errors) ? errors[0] : null;
  return first?.message || "Speechmatics transcription failed";
}

/**
 * Poll a submitted Speechmatics job until it reaches a terminal state
 * (max 120s), then fetch its transcript.
 */
async function pollSpeechmaticsJob(jobUrl, authHeaders) {
  const maxWait = 120_000;
  const start = Date.now();

  while (Date.now() - start < maxWait) {
    await new Promise((r) => setTimeout(r, 2000));

    const pollRes = await fetch(jobUrl, { headers: authHeaders });
    if (!pollRes.ok) continue;

    const result = await pollRes.json();
    const status = result?.job?.status;

    if (status === "done") {
      return fetchSpeechmaticsTranscript(jobUrl, authHeaders);
    }

    if (status === "rejected") {
      return errorResponse(500, speechmaticsJobErrorMessage(result));
    }
  }

  return errorResponse(504, "Speechmatics transcription timed out after 120s");
}

/**
 * Handle Speechmatics transcription (async batch: submit multipart job → poll → fetch transcript)
 *
 * Speechmatics batch mode accepts the audio file directly in the job-submission
 * multipart body (field "data_file") alongside a JSON "config" field describing
 * the requested transcription options. Streaming (real-time WebSocket) mode is
 * out of scope for v1 — this handler only implements batch (REST) transcription.
 */
async function handleSpeechmaticsTranscription(providerConfig, file, modelId, token) {
  const authHeaders = buildAuthHeaders(providerConfig, token);
  const baseUrl = providerConfig.baseUrl.replace(/\/$/, "");

  // Step 1: submit the job — multipart body with "data_file" (audio) + "config" (JSON)
  const config = JSON.stringify({
    type: "transcription",
    transcription_config: { operating_point: speechmaticsOperatingPoint(modelId) },
  });
  const { body, contentType } = await buildMultipartBody(file, { config }, "data_file");

  const submitRes = await fetch(baseUrl, {
    method: "POST",
    headers: { ...authHeaders, "Content-Type": contentType },
    body,
  });

  if (!submitRes.ok) {
    return upstreamErrorResponse(submitRes, await submitRes.text());
  }

  const { id: jobId } = await submitRes.json();
  if (!jobId) {
    return errorResponse(502, "Speechmatics did not return a job id");
  }

  // Step 2: poll for completion (max 120s)
  return pollSpeechmaticsJob(`${baseUrl}/${jobId}`, authHeaders);
}

/**
 * Handle audio transcription request
 *
 * @param {Object} options
 * @param {FormData} options.formData - Multipart form data with file + model
 * @param {Object} options.credentials - Provider credentials { apiKey }
 * @returns {Response}
 */
export async function handleAudioTranscription({
  formData,
  credentials,
  resolvedProvider = null,
  resolvedModel = null,
}: {
  formData: FormData;
  credentials?: TranscriptionCredentials | null;
  resolvedProvider?: AudioProvider | null;
  resolvedModel?: string | null;
}): Promise<Response> {
  const model = formData.get("model");
  if (typeof model !== "string" || !model) {
    return errorResponse(400, "model is required");
  }

  const fileEntry = formData.get("file");
  if (!(fileEntry instanceof Blob)) {
    return errorResponse(400, "file is required");
  }
  const file = fileEntry as Blob & { name?: unknown };

  // Use pre-resolved provider/model from route handler if available (supports dynamic provider_nodes).
  let providerConfig = resolvedProvider;
  let modelId = resolvedModel;
  if (!providerConfig) {
    const parsed = parseTranscriptionModel(model);
    providerConfig = parsed.provider ? getTranscriptionProvider(parsed.provider) : null;
    modelId = parsed.model;
  }

  if (!providerConfig) {
    return errorResponse(
      400,
      `No transcription provider found for model "${model}". Available: openai, openrouter, groq, deepgram, assemblyai, nvidia, huggingface, qwen, gladia, rev-ai, speechmatics`
    );
  }

  // Skip credential check for local providers (authType: "none")
  const token =
    providerConfig.authType === "none" ? null : credentials?.apiKey || credentials?.accessToken;
  if (providerConfig.authType !== "none" && !token) {
    return errorResponse(401, `No credentials for transcription provider: ${providerConfig.id}`);
  }

  // Route to provider-specific handler
  if (providerConfig.format === "vertex-gemini") {
    try {
      const buffer = Buffer.from(await file.arrayBuffer());
      const uploadedType =
        typeof (file as { type?: unknown }).type === "string" && (file as { type?: string }).type
          ? (file as { type: string }).type
          : "audio/wav";
      const languageValue = formData.get("language");
      const promptValue = formData.get("prompt");
      const text = await vertexTranscribe(credentials ?? {}, {
        model: modelId as string,
        audioBase64: buffer.toString("base64"),
        mimeType: uploadedType,
        prompt: typeof promptValue === "string" ? promptValue : undefined,
        language: typeof languageValue === "string" ? languageValue : undefined,
      });
      return Response.json({ text }, { headers: { ...CORS_HEADERS } });
    } catch (err) {
      const error = err as { message?: string; status?: number };
      return errorResponse(
        typeof error?.status === "number" ? error.status : 500,
        `Vertex transcription failed: ${error?.message || "unknown error"}`
      );
    }
  }

  if (providerConfig.format === "deepgram") {
    return handleDeepgramTranscription(providerConfig, file, modelId, token, formData);
  }

  if (providerConfig.format === "assemblyai") {
    return handleAssemblyAITranscription(providerConfig, file, modelId, token);
  }

  if (providerConfig.format === "gladia") {
    return handleGladiaTranscription(providerConfig, file, modelId, token);
  }

  if (providerConfig.format === "nvidia-asr") {
    return handleNvidiaTranscription(providerConfig, file, modelId, token);
  }

  if (providerConfig.format === "huggingface-asr") {
    return handleHuggingFaceTranscription(providerConfig, file, modelId, token);
  }

  if (providerConfig.format === "kie-audio") {
    return handleKieAudioTranscription(providerConfig, file, modelId, token);
  }

  if (providerConfig.format === "rev-ai") {
    return handleRevAiTranscription(providerConfig, file, modelId, token);
  }

  if (providerConfig.format === "speechmatics") {
    return handleSpeechmaticsTranscription(providerConfig, file, modelId, token);
  }

  if (providerConfig.format === "openrouter-stt") {
    return handleOpenRouterTranscription(providerConfig, file, modelId, token, formData);
  }

  // Default: OpenAI/Groq/Qwen3-compatible multipart proxy
  const extraFields: Record<string, string> = {};
  for (const key of [
    "language",
    "prompt",
    "response_format",
    "temperature",
    "timestamp_granularities[]",
  ]) {
    const val = formData.get(key);
    if (val !== null && val !== undefined) {
      extraFields[key] = String(val);
    }
  }

  const { body: multipartBody, contentType: multipartCT } = await buildMultipartBody(file, {
    model: modelId,
    ...extraFields,
  });

  try {
    const res = await fetch(providerConfig.baseUrl, {
      method: "POST",
      headers: { ...buildAuthHeaders(providerConfig, token), "Content-Type": multipartCT },
      body: multipartBody,
    });

    if (!res.ok) {
      return upstreamErrorResponse(res, await res.text());
    }

    const data = await res.text();
    const respContentType = res.headers.get("content-type") || "application/json";

    return new Response(data, {
      status: 200,
      headers: { "Content-Type": respContentType },
    });
  } catch (err) {
    const error = err instanceof Error ? err : new Error(String(err));
    return errorResponse(500, `Transcription request failed: ${error.message}`);
  }
}