@cloudflare/tanstack-ai 0.0.0 → 0.1.1

package/dist/chunk-UUFEOQ6B.js

@@ -0,0 +1,315 @@
+// src/utils/create-fetcher.ts
+function isDirectBindingConfig(config) {
+  return "binding" in config && typeof config.binding.gateway === "function";
+}
+function isDirectCredentialsConfig(config) {
+  return "accountId" in config && "apiKey" in config && !("gatewayId" in config);
+}
+function createGatewayFetch(provider, config, headers = {}) {
+  return (input, init) => {
+    let query = {};
+    const url = typeof input === "string" ? input : input instanceof URL ? input.href : input.url;
+    const urlObj = new URL(url);
+    const endpoint = urlObj.pathname.replace(/^\/v1\//, "").replace(/^\//, "") + urlObj.search;
+    if (init?.body) {
+      try {
+        query = JSON.parse(init.body);
+      } catch {
+        query = { _raw: init.body };
+      }
+    }
+    const cacheHeaders = {};
+    if ("skipCache" in config && config.skipCache) {
+      cacheHeaders["cf-aig-skip-cache"] = "true";
+    }
+    if (typeof config.cacheTtl === "number") {
+      cacheHeaders["cf-aig-cache-ttl"] = String(config.cacheTtl);
+    }
+    if (typeof config.customCacheKey === "string") {
+      cacheHeaders["cf-aig-cache-key"] = config.customCacheKey;
+    }
+    if (typeof config.metadata === "object") {
+      cacheHeaders["cf-aig-metadata"] = JSON.stringify(config.metadata);
+    }
+    const request = {
+      provider,
+      endpoint,
+      headers: {
+        ...init?.headers,
+        ...headers,
+        ...cacheHeaders,
+        "Content-Type": "application/json"
+      },
+      query
+    };
+    if (provider === "workers-ai") {
+      request.endpoint = query.model;
+      delete query.model;
+      delete query.instructions;
+    }
+    if (config.apiKey) {
+      request.headers["authorization"] = `Bearer ${config.apiKey}`;
+    }
+    if ("binding" in config) {
+      return config.binding.run(request);
+    }
+    return fetch(
+      `https://gateway.ai.cloudflare.com/v1/${config.accountId}/${config.gatewayId}`,
+      {
+        ...init,
+        headers: {
+          "Content-Type": "application/json",
+          ...headers,
+          ...cacheHeaders,
+          ...config.cfApiKey ? { "cf-aig-authorization": `Bearer ${config.cfApiKey}` } : {}
+        },
+        body: JSON.stringify(request)
+      }
+    );
+  };
+}
+function normalizeMessagesForBinding(messages) {
+  return messages.map((msg) => {
+    const normalized = { ...msg };
+    if (normalized.content === null || normalized.content === void 0) {
+      normalized.content = "";
+    }
+    if (normalized.tool_call_id && typeof normalized.tool_call_id === "string") {
+      normalized.tool_call_id = sanitizeToolCallId(normalized.tool_call_id);
+    }
+    if (Array.isArray(normalized.tool_calls)) {
+      normalized.tool_calls = normalized.tool_calls.map(
+        (tc) => {
+          if (tc.id && typeof tc.id === "string") {
+            return { ...tc, id: sanitizeToolCallId(tc.id) };
+          }
+          return tc;
+        }
+      );
+    }
+    return normalized;
+  });
+}
+function sanitizeToolCallId(id) {
+  const alphanumeric = id.replace(/[^a-zA-Z0-9]/g, "");
+  return alphanumeric.slice(0, 9).padEnd(9, "0");
+}
+function createWorkersAiBindingFetch(binding) {
+  return async (_input, init) => {
+    if (!init?.body) {
+      return new Response("No body", { status: 400 });
+    }
+    let body;
+    try {
+      body = JSON.parse(init.body);
+    } catch {
+      return new Response("Invalid JSON body", { status: 400 });
+    }
+    const model = body.model;
+    const stream = body.stream;
+    const inputs = {};
+    if (body.messages) {
+      inputs.messages = normalizeMessagesForBinding(
+        body.messages
+      );
+    }
+    if (body.tools) inputs.tools = body.tools;
+    if (typeof body.temperature === "number") inputs.temperature = body.temperature;
+    if (typeof body.max_tokens === "number") inputs.max_tokens = body.max_tokens;
+    if (body.response_format) inputs.response_format = body.response_format;
+    if (stream) inputs.stream = true;
+    const result = await binding.run(model, inputs);
+    if (stream && result instanceof ReadableStream) {
+      const transformed = transformWorkersAiStream(
+        result,
+        model
+      );
+      return new Response(transformed, {
+        headers: {
+          "Content-Type": "text/event-stream",
+          "Cache-Control": "no-cache"
+        }
+      });
+    }
+    const responseObj = typeof result === "object" && result !== null ? result : { response: String(result) };
+    const responseText = typeof responseObj.response === "string" ? responseObj.response : "";
+    const message = {
+      role: "assistant",
+      content: responseText
+    };
+    let finishReason = "stop";
+    if (Array.isArray(responseObj.tool_calls) && responseObj.tool_calls.length > 0) {
+      finishReason = "tool_calls";
+      message.tool_calls = responseObj.tool_calls.map(
+        (tc) => ({
+          id: sanitizeToolCallId(tc.id || crypto.randomUUID()),
+          type: "function",
+          function: {
+            name: tc.function?.name || tc.name || "",
+            arguments: typeof (tc.function?.arguments ?? tc.arguments) === "string" ? tc.function?.arguments ?? tc.arguments : JSON.stringify(tc.function?.arguments ?? tc.arguments ?? {})
+          }
+        })
+      );
+    }
+    const openAiResponse = {
+      id: `workers-ai-${crypto.randomUUID()}`,
+      object: "chat.completion",
+      created: Math.floor(Date.now() / 1e3),
+      model,
+      choices: [{ index: 0, message, finish_reason: finishReason }]
+    };
+    return new Response(JSON.stringify(openAiResponse), {
+      headers: { "Content-Type": "application/json" }
+    });
+  };
+}
+function transformWorkersAiStream(source, model) {
+  const decoder = new TextDecoder();
+  const encoder = new TextEncoder();
+  const streamId = `workers-ai-${crypto.randomUUID()}`;
+  const created = Math.floor(Date.now() / 1e3);
+  let buffer = "";
+  let hasToolCalls = false;
+  let isOpenAiFormat = false;
+  const emittedToolCallStart = /* @__PURE__ */ new Set();
+  return source.pipeThrough(
+    new TransformStream({
+      transform(chunk, controller) {
+        buffer += decoder.decode(chunk, { stream: true });
+        const lines = buffer.split("\n");
+        buffer = lines.pop() || "";
+        for (const line of lines) {
+          const trimmed = line.trim();
+          if (!trimmed || !trimmed.startsWith("data: ")) continue;
+          const data = trimmed.slice(6);
+          if (data === "[DONE]") continue;
+          try {
+            const parsed = JSON.parse(data);
+            if (parsed.choices !== void 0) {
+              isOpenAiFormat = true;
+              const choice = parsed.choices?.[0];
+              if (choice?.delta?.tool_calls) {
+                hasToolCalls = true;
+                for (const tc of choice.delta.tool_calls) {
+                  if (tc.id && typeof tc.id === "string") {
+                    tc.id = sanitizeToolCallId(tc.id);
+                  }
+                }
+              }
+              if (choice?.finish_reason === "tool_calls") {
+                hasToolCalls = true;
+              }
+              controller.enqueue(
+                encoder.encode(`data: ${JSON.stringify(parsed)}
+
+`)
+              );
+              continue;
+            }
+            if (parsed.response != null && parsed.response !== "") {
+              const openAiChunk = {
+                id: streamId,
+                object: "chat.completion.chunk",
+                created,
+                model,
+                choices: [
+                  {
+                    index: 0,
+                    delta: { content: parsed.response },
+                    finish_reason: null
+                  }
+                ]
+              };
+              controller.enqueue(
+                encoder.encode(`data: ${JSON.stringify(openAiChunk)}
+
+`)
+              );
+            }
+            if (Array.isArray(parsed.tool_calls) && parsed.tool_calls.length > 0) {
+              for (const tc of parsed.tool_calls) {
+                const tcIndex = tc.index ?? 0;
+                const tcName = tc.function?.name ?? tc.name ?? null;
+                const tcArgs = tc.function?.arguments ?? tc.arguments ?? null;
+                const tcId = tc.id ?? null;
+                if (!tcId && !tcName && (!tcArgs || tcArgs === "")) continue;
+                hasToolCalls = true;
+                const toolCallDelta = {
+                  index: tcIndex
+                };
+                if (!emittedToolCallStart.has(tcIndex)) {
+                  emittedToolCallStart.add(tcIndex);
+                  const rawId = tcId || `call${streamId}${tcIndex}`;
+                  toolCallDelta.id = sanitizeToolCallId(rawId);
+                  toolCallDelta.type = "function";
+                  toolCallDelta.function = {
+                    name: tcName || "",
+                    // Include arguments if they arrive in the same chunk
+                    arguments: tcArgs != null ? typeof tcArgs === "string" ? tcArgs : JSON.stringify(tcArgs) : ""
+                  };
+                } else {
+                  if (tcArgs != null && tcArgs !== "") {
+                    toolCallDelta.function = {
+                      arguments: typeof tcArgs === "string" ? tcArgs : JSON.stringify(tcArgs)
+                    };
+                  } else {
+                    continue;
+                  }
+                }
+                const toolChunk = {
+                  id: streamId,
+                  object: "chat.completion.chunk",
+                  created,
+                  model,
+                  choices: [
+                    {
+                      index: 0,
+                      delta: { tool_calls: [toolCallDelta] },
+                      finish_reason: null
+                    }
+                  ]
+                };
+                controller.enqueue(
+                  encoder.encode(`data: ${JSON.stringify(toolChunk)}
+
+`)
+                );
+              }
+            }
+          } catch (e) {
+            console.warn("[tanstack-ai] failed to parse SSE event:", data, e);
+          }
+        }
+      },
+      flush(controller) {
+        if (!isOpenAiFormat) {
+          const finalChunk = {
+            id: streamId,
+            object: "chat.completion.chunk",
+            created,
+            model,
+            choices: [
+              {
+                index: 0,
+                delta: {},
+                finish_reason: hasToolCalls ? "tool_calls" : "stop"
+              }
+            ]
+          };
+          controller.enqueue(encoder.encode(`data: ${JSON.stringify(finalChunk)}
+
+`));
+        }
+        controller.enqueue(encoder.encode("data: [DONE]\n\n"));
+      }
+    })
+  );
+}
+
+export {
+  isDirectBindingConfig,
+  isDirectCredentialsConfig,
+  createGatewayFetch,
+  createWorkersAiBindingFetch
+};
+//# sourceMappingURL=chunk-UUFEOQ6B.js.map

package/dist/chunk-UUFEOQ6B.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/utils/create-fetcher.ts"],"sourcesContent":["// ---------------------------------------------------------------------------\n// AI Gateway types (for third-party providers + Workers AI through gateway)\n// ---------------------------------------------------------------------------\n\nexport interface CloudflareAiGateway {\n\trun(request: unknown): Promise<Response>;\n}\n\nexport interface AiGatewayBindingConfig {\n\t/**\n\t * The AI Gateway binding\n\t * @example\n\t * env.AI.gateway('my-gateway-id')\n\t */\n\tbinding: CloudflareAiGateway;\n\t/**\n\t * The Provider API Key if you want to manually pass it, ignore if using Unified Billing or BYOK.\n\t */\n\tapiKey?: string;\n}\n\nexport interface AiGatewayCredentialsConfig {\n\t/**\n\t * The Cloudflare account ID\n\t */\n\taccountId: string;\n\t/**\n\t * The AI Gateway ID\n\t */\n\tgatewayId: string;\n\t/**\n\t * The Provider API Key if you want to manually pass it, ignore if using Unified Billing or BYOK.\n\t */\n\tapiKey?: string;\n\t/**\n\t * The Cloudflare AI Gateway API Key, required if your Gateway is authenticated.\n\t */\n\tcfApiKey?: string;\n}\n\nexport interface AiGatewayConfig {\n\tskipCache?: boolean;\n\tcacheTtl?: number;\n\tcustomCacheKey?: string;\n\tmetadata?: Record<string, unknown>;\n}\n\nexport type AiGatewayAdapterConfig = (AiGatewayBindingConfig | AiGatewayCredentialsConfig) &\n\tAiGatewayConfig;\n\n// ---------------------------------------------------------------------------\n// Plain Workers AI types (direct binding or REST, no gateway)\n// ---------------------------------------------------------------------------\n\n/**\n * The Workers AI binding interface (env.AI).\n * Accepts a model name and inputs, returns results directly.\n * Includes `gateway()` which is present on `env.AI` but not on `env.AI.gateway(id)`,\n * enabling structural discrimination from `CloudflareAiGateway`.\n */\nexport interface WorkersAiBinding {\n\trun(\n\t\tmodel: string,\n\t\tinputs: Record<string, unknown>,\n\t\toptions?: Record<string, unknown>,\n\t): Promise<unknown>;\n\tgateway(gatewayId: string): CloudflareAiGateway;\n}\n\nexport interface WorkersAiDirectBindingConfig {\n\t/**\n\t * The Workers AI binding (env.AI).\n\t * @example\n\t * { binding: env.AI }\n\t */\n\tbinding: WorkersAiBinding;\n}\n\nexport interface WorkersAiDirectCredentialsConfig {\n\t/**\n\t * The Cloudflare account ID\n\t */\n\taccountId: string;\n\t/**\n\t * The Cloudflare API key for Workers AI\n\t */\n\tapiKey: string;\n}\n\n/**\n * Config for Workers AI adapters. Supports four modes:\n * - Plain binding: `{ binding: env.AI }`\n * - Plain REST: `{ accountId, apiKey }`\n * - AI Gateway binding: `{ binding: env.AI.gateway(id) }`\n * - AI Gateway REST: `{ accountId, gatewayId, ... }`\n *\n * The third union member intersects `AiGatewayAdapterConfig` with `{ apiKey?: string }`.\n * For the gateway binding variant, `AiGatewayBindingConfig` already includes `apiKey?`,\n * so the intersection is redundant there. For the gateway credentials variant, this\n * `apiKey` represents the Workers AI token (used in the `Authorization` header to the\n * upstream provider), distinct from `cfApiKey` (used in the `cf-aig-authorization`\n * header for authenticated gateways).\n */\nexport type WorkersAiAdapterConfig =\n\t| WorkersAiDirectBindingConfig\n\t| WorkersAiDirectCredentialsConfig\n\t| (AiGatewayAdapterConfig & { apiKey?: string });\n\n// ---------------------------------------------------------------------------\n// Config detection helpers\n// ---------------------------------------------------------------------------\n\n/** Returns true if this is a plain Workers AI binding config (`{ binding: env.AI }`) */\nexport function isDirectBindingConfig(\n\tconfig: WorkersAiAdapterConfig,\n): config is WorkersAiDirectBindingConfig {\n\t// env.AI has a .gateway() method; env.AI.gateway(id) does not.\n\t// This distinguishes direct bindings from AI Gateway bindings.\n\treturn (\n\t\t\"binding\" in config &&\n\t\ttypeof (config.binding as unknown as Record<string, unknown>).gateway === \"function\"\n\t);\n}\n\n/** Returns true if this is a plain Workers AI REST config (accountId + apiKey, no gatewayId) */\nexport function isDirectCredentialsConfig(\n\tconfig: WorkersAiAdapterConfig,\n): config is WorkersAiDirectCredentialsConfig {\n\treturn \"accountId\" in config && \"apiKey\" in config && !(\"gatewayId\" in config);\n}\n\n/** Returns true if this is an AI Gateway config (has gateway binding or `gatewayId`) */\nexport function isGatewayConfig(config: WorkersAiAdapterConfig): config is AiGatewayAdapterConfig {\n\tif (\"gatewayId\" in config) return true;\n\t// Has `binding` but NOT a direct Workers AI binding (no .gateway method)\n\treturn \"binding\" in config && !isDirectBindingConfig(config);\n}\n\n// ---------------------------------------------------------------------------\n// createGatewayFetch -- for routing through AI Gateway\n// ---------------------------------------------------------------------------\n\nexport function createGatewayFetch(\n\tprovider: string,\n\tconfig: AiGatewayAdapterConfig,\n\theaders: Record<string, string> = {},\n): typeof fetch {\n\treturn (input, init) => {\n\t\tlet query: Record<string, unknown> = {};\n\n\t\tconst url =\n\t\t\ttypeof input === \"string\" ? input : input instanceof URL ? input.href : input.url;\n\t\tconst urlObj = new URL(url);\n\n\t\t// Extract endpoint path (remove /v1/ prefix if present)\n\t\tconst endpoint = urlObj.pathname.replace(/^\\/v1\\//, \"\").replace(/^\\//, \"\") + urlObj.search;\n\n\t\tif (init?.body) {\n\t\t\ttry {\n\t\t\t\tquery = JSON.parse(init.body as string);\n\t\t\t} catch {\n\t\t\t\tquery = { _raw: init.body };\n\t\t\t}\n\t\t}\n\n\t\tconst cacheHeaders: Record<string, string> = {};\n\n\t\tif (\"skipCache\" in config && config.skipCache) {\n\t\t\tcacheHeaders[\"cf-aig-skip-cache\"] = \"true\";\n\t\t}\n\n\t\tif (typeof config.cacheTtl === \"number\") {\n\t\t\tcacheHeaders[\"cf-aig-cache-ttl\"] = String(config.cacheTtl);\n\t\t}\n\n\t\tif (typeof config.customCacheKey === \"string\") {\n\t\t\tcacheHeaders[\"cf-aig-cache-key\"] = config.customCacheKey;\n\t\t}\n\n\t\tif (typeof config.metadata === \"object\") {\n\t\t\tcacheHeaders[\"cf-aig-metadata\"] = JSON.stringify(config.metadata);\n\t\t}\n\n\t\tconst request: {\n\t\t\tprovider: string;\n\t\t\tendpoint: string;\n\t\t\theaders: Record<string, string>;\n\t\t\tquery: Record<string, unknown>;\n\t\t} = {\n\t\t\tprovider,\n\t\t\tendpoint,\n\t\t\theaders: {\n\t\t\t\t...(init?.headers as Record<string, string> | undefined),\n\t\t\t\t...headers,\n\t\t\t\t...cacheHeaders,\n\t\t\t\t\"Content-Type\": \"application/json\",\n\t\t\t},\n\t\t\tquery,\n\t\t};\n\n\t\tif (provider === \"workers-ai\") {\n\t\t\trequest.endpoint = query.model as string;\n\t\t\tdelete query.model;\n\t\t\tdelete query.instructions;\n\t\t}\n\n\t\tif (config.apiKey) {\n\t\t\trequest.headers[\"authorization\"] = `Bearer ${config.apiKey}`;\n\t\t}\n\n\t\tif (\"binding\" in config) {\n\t\t\treturn config.binding.run(request);\n\t\t}\n\n\t\treturn fetch(\n\t\t\t`https://gateway.ai.cloudflare.com/v1/${config.accountId}/${config.gatewayId}`,\n\t\t\t{\n\t\t\t\t...init,\n\t\t\t\theaders: {\n\t\t\t\t\t\"Content-Type\": \"application/json\",\n\t\t\t\t\t...headers,\n\t\t\t\t\t...cacheHeaders,\n\t\t\t\t\t...(config.cfApiKey\n\t\t\t\t\t\t? { \"cf-aig-authorization\": `Bearer ${config.cfApiKey}` }\n\t\t\t\t\t\t: {}),\n\t\t\t\t},\n\t\t\t\tbody: JSON.stringify(request),\n\t\t\t},\n\t\t);\n\t};\n}\n\n// ---------------------------------------------------------------------------\n// createWorkersAiBindingFetch -- shim that makes env.AI look like an OpenAI endpoint\n// ---------------------------------------------------------------------------\n\n/**\n * Normalize messages before passing to Workers AI binding.\n *\n * The binding has strict schema validation that may differ from the OpenAI API:\n * - `content` must be a string (not null)\n * - `tool_call_id` must match `[a-zA-Z0-9]{9}` pattern\n *\n * This function patches these fields so that the full tool-call round-trip works\n * even though the binding's own generated IDs may not pass its validation.\n */\nfunction normalizeMessagesForBinding(\n\tmessages: Record<string, unknown>[],\n): Record<string, unknown>[] {\n\treturn messages.map((msg) => {\n\t\tconst normalized = { ...msg };\n\n\t\t// content: null → content: \"\"\n\t\tif (normalized.content === null || normalized.content === undefined) {\n\t\t\tnormalized.content = \"\";\n\t\t}\n\n\t\t// Normalize tool_call_id on tool messages\n\t\tif (normalized.tool_call_id && typeof normalized.tool_call_id === \"string\") {\n\t\t\tnormalized.tool_call_id = sanitizeToolCallId(normalized.tool_call_id);\n\t\t}\n\n\t\t// Normalize tool_calls[].id on assistant messages\n\t\tif (Array.isArray(normalized.tool_calls)) {\n\t\t\tnormalized.tool_calls = (normalized.tool_calls as Record<string, unknown>[]).map(\n\t\t\t\t(tc) => {\n\t\t\t\t\tif (tc.id && typeof tc.id === \"string\") {\n\t\t\t\t\t\treturn { ...tc, id: sanitizeToolCallId(tc.id) };\n\t\t\t\t\t}\n\t\t\t\t\treturn tc;\n\t\t\t\t},\n\t\t\t);\n\t\t}\n\n\t\treturn normalized;\n\t});\n}\n\n/**\n * Strip non-alphanumeric characters and ensure the ID is exactly 9 chars,\n * matching Workers AI's `[a-zA-Z0-9]{9}` validation pattern.\n *\n * **Why this exists:** The Workers AI binding validates `tool_call_id` with\n * a strict `[a-zA-Z0-9]{9}` regex, but it *generates* IDs like\n * `chatcmpl-tool-875d3ec6179676ae` (with dashes, >9 chars). Those IDs are\n * then rejected when sent back in a follow-up request. This is a known\n * Workers AI issue — see workers-ai.md (Issue 3). Once the Workers AI team\n * fixes the validation, this function becomes an idempotent no-op for\n * IDs that already match the pattern.\n */\nfunction sanitizeToolCallId(id: string): string {\n\tconst alphanumeric = id.replace(/[^a-zA-Z0-9]/g, \"\");\n\t// Pad with zeros if too short, truncate if too long\n\treturn alphanumeric.slice(0, 9).padEnd(9, \"0\");\n}\n\n/**\n * Creates a fetch function that intercepts OpenAI SDK requests and translates them\n * to Workers AI binding calls (env.AI.run). This allows the WorkersAiTextAdapter\n * to use the OpenAI SDK against a plain Workers AI binding.\n *\n * NOTE: The `input` URL parameter is intentionally ignored. The model name and all\n * request parameters are extracted from the JSON body, matching Workers AI's\n * `binding.run(model, inputs)` calling convention.\n */\nexport function createWorkersAiBindingFetch(binding: WorkersAiBinding): typeof fetch {\n\treturn async (_input, init) => {\n\t\tif (!init?.body) {\n\t\t\treturn new Response(\"No body\", { status: 400 });\n\t\t}\n\n\t\tlet body: Record<string, unknown>;\n\t\ttry {\n\t\t\tbody = JSON.parse(init.body as string);\n\t\t} catch {\n\t\t\treturn new Response(\"Invalid JSON body\", { status: 400 });\n\t\t}\n\n\t\tconst model = body.model as string;\n\t\tconst stream = body.stream as boolean | undefined;\n\n\t\t// Build Workers AI inputs from OpenAI format\n\t\tconst inputs: Record<string, unknown> = {};\n\t\tif (body.messages) {\n\t\t\tinputs.messages = normalizeMessagesForBinding(\n\t\t\t\tbody.messages as Record<string, unknown>[],\n\t\t\t);\n\t\t}\n\t\tif (body.tools) inputs.tools = body.tools;\n\t\tif (typeof body.temperature === \"number\") inputs.temperature = body.temperature;\n\t\tif (typeof body.max_tokens === \"number\") inputs.max_tokens = body.max_tokens;\n\t\tif (body.response_format) inputs.response_format = body.response_format;\n\t\tif (stream) inputs.stream = true;\n\n\t\tconst result = await binding.run(model, inputs);\n\n\t\tif (stream && result instanceof ReadableStream) {\n\t\t\t// Workers AI returns an SSE stream with `data: {\"response\":\"chunk\"}` format.\n\t\t\t// Transform it to OpenAI-compatible SSE format.\n\t\t\tconst transformed = transformWorkersAiStream(\n\t\t\t\tresult as ReadableStream<Uint8Array>,\n\t\t\t\tmodel,\n\t\t\t);\n\t\t\treturn new Response(transformed, {\n\t\t\t\theaders: {\n\t\t\t\t\t\"Content-Type\": \"text/event-stream\",\n\t\t\t\t\t\"Cache-Control\": \"no-cache\",\n\t\t\t\t},\n\t\t\t});\n\t\t}\n\n\t\t// Graceful degradation: some models return a complete (non-streaming)\n\t\t// response even when `stream: true` is requested. Fall through to the\n\t\t// non-streaming wrapper which produces a valid OpenAI Chat Completion\n\t\t// response that the SDK can consume.\n\n\t\t// Non-streaming: Workers AI returns { response: \"text\", tool_calls?: [...] }\n\t\t// Wrap into OpenAI Chat Completion format.\n\t\tconst responseObj =\n\t\t\ttypeof result === \"object\" && result !== null\n\t\t\t\t? (result as Record<string, unknown>)\n\t\t\t\t: { response: String(result) };\n\n\t\tconst responseText = typeof responseObj.response === \"string\" ? responseObj.response : \"\";\n\n\t\tconst message: Record<string, unknown> = {\n\t\t\trole: \"assistant\",\n\t\t\tcontent: responseText,\n\t\t};\n\t\tlet finishReason = \"stop\";\n\n\t\t// Handle tool calls if present in Workers AI response\n\t\tif (Array.isArray(responseObj.tool_calls) && responseObj.tool_calls.length > 0) {\n\t\t\tfinishReason = \"tool_calls\";\n\t\t\tmessage.tool_calls = responseObj.tool_calls.map(\n\t\t\t\t(tc: {\n\t\t\t\t\tid?: string;\n\t\t\t\t\tname?: string;\n\t\t\t\t\targuments: unknown;\n\t\t\t\t\tfunction?: { name: string; arguments?: unknown };\n\t\t\t\t}) => ({\n\t\t\t\t\tid: sanitizeToolCallId(tc.id || crypto.randomUUID()),\n\t\t\t\t\ttype: \"function\",\n\t\t\t\t\tfunction: {\n\t\t\t\t\t\tname: tc.function?.name || tc.name || \"\",\n\t\t\t\t\t\targuments:\n\t\t\t\t\t\t\ttypeof (tc.function?.arguments ?? tc.arguments) === \"string\"\n\t\t\t\t\t\t\t\t? ((tc.function?.arguments ?? tc.arguments) as string)\n\t\t\t\t\t\t\t\t: JSON.stringify(tc.function?.arguments ?? tc.arguments ?? {}),\n\t\t\t\t\t},\n\t\t\t\t}),\n\t\t\t);\n\t\t}\n\n\t\tconst openAiResponse = {\n\t\t\tid: `workers-ai-${crypto.randomUUID()}`,\n\t\t\tobject: \"chat.completion\",\n\t\t\tcreated: Math.floor(Date.now() / 1000),\n\t\t\tmodel,\n\t\t\tchoices: [{ index: 0, message, finish_reason: finishReason }],\n\t\t};\n\n\t\treturn new Response(JSON.stringify(openAiResponse), {\n\t\t\theaders: { \"Content-Type\": \"application/json\" },\n\t\t});\n\t};\n}\n\n// ---------------------------------------------------------------------------\n// Stream transformer: Workers AI SSE -> OpenAI-compatible SSE\n// Uses TransformStream for proper backpressure.\n// ---------------------------------------------------------------------------\n\n/**\n * Transforms a Workers AI SSE stream (data: {\"response\":\"chunk\"}) into\n * an OpenAI-compatible SSE stream (data: {\"choices\":[{\"delta\":{\"content\":\"chunk\"}}]}).\n *\n * Workers AI binding streams tool calls in an OpenAI-like nested format:\n *   { tool_calls: [{ id, type, index, function: { name, arguments } }] }\n * Arguments are streamed incrementally across multiple SSE chunks, so the\n * transformer must forward them as incremental deltas rather than a single blob.\n */\nfunction transformWorkersAiStream(\n\tsource: ReadableStream<Uint8Array>,\n\tmodel: string,\n): ReadableStream<Uint8Array> {\n\tconst decoder = new TextDecoder();\n\tconst encoder = new TextEncoder();\n\t// Generate a stable ID and timestamp for the entire stream, matching OpenAI's\n\t// convention where all chunks in a single response share the same id/created.\n\tconst streamId = `workers-ai-${crypto.randomUUID()}`;\n\tconst created = Math.floor(Date.now() / 1000);\n\tlet buffer = \"\";\n\tlet hasToolCalls = false;\n\t// When true, the source stream is already in OpenAI format (some models\n\t// like Qwen3, Kimi K2.5 stream OpenAI-compatible SSE through the binding).\n\t// In that case, flush() should only emit [DONE] and skip the finish chunk.\n\tlet isOpenAiFormat = false;\n\t// Track which tool call indices we've already emitted an `id` for,\n\t// so subsequent argument deltas don't duplicate the id/type/name fields.\n\tconst emittedToolCallStart = new Set<number>();\n\n\treturn source.pipeThrough(\n\t\tnew TransformStream<Uint8Array, Uint8Array>({\n\t\t\ttransform(chunk, controller) {\n\t\t\t\tbuffer += decoder.decode(chunk, { stream: true });\n\t\t\t\tconst lines = buffer.split(\"\\n\");\n\t\t\t\tbuffer = lines.pop() || \"\";\n\n\t\t\t\tfor (const line of lines) {\n\t\t\t\t\tconst trimmed = line.trim();\n\t\t\t\t\tif (!trimmed || !trimmed.startsWith(\"data: \")) continue;\n\t\t\t\t\tconst data = trimmed.slice(6);\n\n\t\t\t\t\t// Swallow source [DONE]; we emit our own in flush()\n\t\t\t\t\tif (data === \"[DONE]\") continue;\n\n\t\t\t\t\ttry {\n\t\t\t\t\t\tconst parsed = JSON.parse(data);\n\n\t\t\t\t\t\t// Some models (Qwen3, Kimi K2.5) return OpenAI-compatible format\n\t\t\t\t\t\t// directly through the binding, with `choices[].delta.content` and\n\t\t\t\t\t\t// optional `reasoning_content`. Detect this and pass through as-is.\n\t\t\t\t\t\tif (parsed.choices !== undefined) {\n\t\t\t\t\t\t\t// Already OpenAI format — pass through with only tool_call_id\n\t\t\t\t\t\t\t// sanitization for any tool calls present.\n\t\t\t\t\t\t\tisOpenAiFormat = true;\n\t\t\t\t\t\t\tconst choice = parsed.choices?.[0];\n\t\t\t\t\t\t\tif (choice?.delta?.tool_calls) {\n\t\t\t\t\t\t\t\thasToolCalls = true;\n\t\t\t\t\t\t\t\tfor (const tc of choice.delta.tool_calls) {\n\t\t\t\t\t\t\t\t\tif (tc.id && typeof tc.id === \"string\") {\n\t\t\t\t\t\t\t\t\t\ttc.id = sanitizeToolCallId(tc.id);\n\t\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\tif (choice?.finish_reason === \"tool_calls\") {\n\t\t\t\t\t\t\t\thasToolCalls = true;\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\tcontroller.enqueue(\n\t\t\t\t\t\t\t\tencoder.encode(`data: ${JSON.stringify(parsed)}\\n\\n`),\n\t\t\t\t\t\t\t);\n\t\t\t\t\t\t\tcontinue;\n\t\t\t\t\t\t}\n\n\t\t\t\t\t\t// --- Workers AI native format handling below ---\n\n\t\t\t\t\t\t// Text content\n\t\t\t\t\t\tif (parsed.response != null && parsed.response !== \"\") {\n\t\t\t\t\t\t\tconst openAiChunk = {\n\t\t\t\t\t\t\t\tid: streamId,\n\t\t\t\t\t\t\t\tobject: \"chat.completion.chunk\",\n\t\t\t\t\t\t\t\tcreated,\n\t\t\t\t\t\t\t\tmodel,\n\t\t\t\t\t\t\t\tchoices: [\n\t\t\t\t\t\t\t\t\t{\n\t\t\t\t\t\t\t\t\t\tindex: 0,\n\t\t\t\t\t\t\t\t\t\tdelta: { content: parsed.response },\n\t\t\t\t\t\t\t\t\t\tfinish_reason: null,\n\t\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t\t],\n\t\t\t\t\t\t\t};\n\t\t\t\t\t\t\tcontroller.enqueue(\n\t\t\t\t\t\t\t\tencoder.encode(`data: ${JSON.stringify(openAiChunk)}\\n\\n`),\n\t\t\t\t\t\t\t);\n\t\t\t\t\t\t}\n\n\t\t\t\t\t\t// Tool calls — Workers AI binding streams these incrementally:\n\t\t\t\t\t\t//   Chunk A: { id, type, index, function: { name } }          — start\n\t\t\t\t\t\t//   Chunk B: { index, function: { arguments: \"partial...\" } }  — args delta\n\t\t\t\t\t\t//   Chunk C: { index, function: { arguments: \"rest...\" } }     — args delta\n\t\t\t\t\t\t//   Chunk D: { id: null, type: null, index, function: { name: null, arguments: \"\" } } — finalize (skip)\n\t\t\t\t\t\tif (Array.isArray(parsed.tool_calls) && parsed.tool_calls.length > 0) {\n\t\t\t\t\t\t\tfor (const tc of parsed.tool_calls) {\n\t\t\t\t\t\t\t\tconst tcIndex = tc.index ?? 0;\n\n\t\t\t\t\t\t\t\t// Resolve name and arguments from either nested or flat format\n\t\t\t\t\t\t\t\tconst tcName = tc.function?.name ?? tc.name ?? null;\n\t\t\t\t\t\t\t\tconst tcArgs = tc.function?.arguments ?? tc.arguments ?? null;\n\t\t\t\t\t\t\t\tconst tcId = tc.id ?? null;\n\n\t\t\t\t\t\t\t\t// Skip finalization chunks where everything is null/empty\n\t\t\t\t\t\t\t\tif (!tcId && !tcName && (!tcArgs || tcArgs === \"\")) continue;\n\n\t\t\t\t\t\t\t\thasToolCalls = true;\n\n\t\t\t\t\t\t\t\t// Build the OpenAI-compatible tool_calls delta\n\t\t\t\t\t\t\t\tconst toolCallDelta: Record<string, unknown> = {\n\t\t\t\t\t\t\t\t\tindex: tcIndex,\n\t\t\t\t\t\t\t\t};\n\n\t\t\t\t\t\t\t\tif (!emittedToolCallStart.has(tcIndex)) {\n\t\t\t\t\t\t\t\t\t// First chunk for this tool call index — emit id, type, name.\n\t\t\t\t\t\t\t\t\t// Use sanitizeToolCallId so the ID survives round-trip through\n\t\t\t\t\t\t\t\t\t// the binding's strict `[a-zA-Z0-9]{9}` validation.\n\t\t\t\t\t\t\t\t\temittedToolCallStart.add(tcIndex);\n\t\t\t\t\t\t\t\t\tconst rawId = tcId || `call${streamId}${tcIndex}`;\n\t\t\t\t\t\t\t\t\ttoolCallDelta.id = sanitizeToolCallId(rawId);\n\t\t\t\t\t\t\t\t\ttoolCallDelta.type = \"function\";\n\t\t\t\t\t\t\t\t\ttoolCallDelta.function = {\n\t\t\t\t\t\t\t\t\t\tname: tcName || \"\",\n\t\t\t\t\t\t\t\t\t\t// Include arguments if they arrive in the same chunk\n\t\t\t\t\t\t\t\t\t\targuments:\n\t\t\t\t\t\t\t\t\t\t\ttcArgs != null\n\t\t\t\t\t\t\t\t\t\t\t\t? typeof tcArgs === \"string\"\n\t\t\t\t\t\t\t\t\t\t\t\t\t? tcArgs\n\t\t\t\t\t\t\t\t\t\t\t\t\t: JSON.stringify(tcArgs)\n\t\t\t\t\t\t\t\t\t\t\t\t: \"\",\n\t\t\t\t\t\t\t\t\t};\n\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\t// Subsequent chunks — only include arguments delta\n\t\t\t\t\t\t\t\t\tif (tcArgs != null && tcArgs !== \"\") {\n\t\t\t\t\t\t\t\t\t\ttoolCallDelta.function = {\n\t\t\t\t\t\t\t\t\t\t\targuments:\n\t\t\t\t\t\t\t\t\t\t\t\ttypeof tcArgs === \"string\"\n\t\t\t\t\t\t\t\t\t\t\t\t\t? tcArgs\n\t\t\t\t\t\t\t\t\t\t\t\t\t: JSON.stringify(tcArgs),\n\t\t\t\t\t\t\t\t\t\t};\n\t\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\t\tcontinue; // Nothing useful to forward\n\t\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t}\n\n\t\t\t\t\t\t\t\tconst toolChunk = {\n\t\t\t\t\t\t\t\t\tid: streamId,\n\t\t\t\t\t\t\t\t\tobject: \"chat.completion.chunk\",\n\t\t\t\t\t\t\t\t\tcreated,\n\t\t\t\t\t\t\t\t\tmodel,\n\t\t\t\t\t\t\t\t\tchoices: [\n\t\t\t\t\t\t\t\t\t\t{\n\t\t\t\t\t\t\t\t\t\t\tindex: 0,\n\t\t\t\t\t\t\t\t\t\t\tdelta: { tool_calls: [toolCallDelta] },\n\t\t\t\t\t\t\t\t\t\t\tfinish_reason: null,\n\t\t\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t\t\t],\n\t\t\t\t\t\t\t\t};\n\t\t\t\t\t\t\t\tcontroller.enqueue(\n\t\t\t\t\t\t\t\t\tencoder.encode(`data: ${JSON.stringify(toolChunk)}\\n\\n`),\n\t\t\t\t\t\t\t\t);\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t}\n\t\t\t\t\t} catch (e) {\n\t\t\t\t\t\t// Log malformed SSE events for debugging; don't break the stream.\n\t\t\t\t\t\tconsole.warn(\"[tanstack-ai] failed to parse SSE event:\", data, e);\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t},\n\t\t\tflush(controller) {\n\t\t\t\tif (!isOpenAiFormat) {\n\t\t\t\t\t// Workers AI native format: emit a finish chunk with stop/tool_calls\n\t\t\t\t\tconst finalChunk = {\n\t\t\t\t\t\tid: streamId,\n\t\t\t\t\t\tobject: \"chat.completion.chunk\",\n\t\t\t\t\t\tcreated,\n\t\t\t\t\t\tmodel,\n\t\t\t\t\t\tchoices: [\n\t\t\t\t\t\t\t{\n\t\t\t\t\t\t\t\tindex: 0,\n\t\t\t\t\t\t\t\tdelta: {},\n\t\t\t\t\t\t\t\tfinish_reason: hasToolCalls ? \"tool_calls\" : \"stop\",\n\t\t\t\t\t\t\t},\n\t\t\t\t\t\t],\n\t\t\t\t\t};\n\t\t\t\t\tcontroller.enqueue(encoder.encode(`data: ${JSON.stringify(finalChunk)}\\n\\n`));\n\t\t\t\t}\n\t\t\t\t// OpenAI format already includes its own finish_reason in the stream.\n\t\t\t\t// Either way, emit a [DONE] sentinel.\n\t\t\t\tcontroller.enqueue(encoder.encode(\"data: [DONE]\\n\\n\"));\n\t\t\t},\n\t\t}),\n\t);\n}\n"],"mappings":";AAiHO,SAAS,sBACf,QACyC;AAGzC,SACC,aAAa,UACb,OAAQ,OAAO,QAA+C,YAAY;AAE5E;AAGO,SAAS,0BACf,QAC6C;AAC7C,SAAO,eAAe,UAAU,YAAY,UAAU,EAAE,eAAe;AACxE;AAaO,SAAS,mBACf,UACA,QACA,UAAkC,CAAC,GACpB;AACf,SAAO,CAAC,OAAO,SAAS;AACvB,QAAI,QAAiC,CAAC;AAEtC,UAAM,MACL,OAAO,UAAU,WAAW,QAAQ,iBAAiB,MAAM,MAAM,OAAO,MAAM;AAC/E,UAAM,SAAS,IAAI,IAAI,GAAG;AAG1B,UAAM,WAAW,OAAO,SAAS,QAAQ,WAAW,EAAE,EAAE,QAAQ,OAAO,EAAE,IAAI,OAAO;AAEpF,QAAI,MAAM,MAAM;AACf,UAAI;AACH,gBAAQ,KAAK,MAAM,KAAK,IAAc;AAAA,MACvC,QAAQ;AACP,gBAAQ,EAAE,MAAM,KAAK,KAAK;AAAA,MAC3B;AAAA,IACD;AAEA,UAAM,eAAuC,CAAC;AAE9C,QAAI,eAAe,UAAU,OAAO,WAAW;AAC9C,mBAAa,mBAAmB,IAAI;AAAA,IACrC;AAEA,QAAI,OAAO,OAAO,aAAa,UAAU;AACxC,mBAAa,kBAAkB,IAAI,OAAO,OAAO,QAAQ;AAAA,IAC1D;AAEA,QAAI,OAAO,OAAO,mBAAmB,UAAU;AAC9C,mBAAa,kBAAkB,IAAI,OAAO;AAAA,IAC3C;AAEA,QAAI,OAAO,OAAO,aAAa,UAAU;AACxC,mBAAa,iBAAiB,IAAI,KAAK,UAAU,OAAO,QAAQ;AAAA,IACjE;AAEA,UAAM,UAKF;AAAA,MACH;AAAA,MACA;AAAA,MACA,SAAS;AAAA,QACR,GAAI,MAAM;AAAA,QACV,GAAG;AAAA,QACH,GAAG;AAAA,QACH,gBAAgB;AAAA,MACjB;AAAA,MACA;AAAA,IACD;AAEA,QAAI,aAAa,cAAc;AAC9B,cAAQ,WAAW,MAAM;AACzB,aAAO,MAAM;AACb,aAAO,MAAM;AAAA,IACd;AAEA,QAAI,OAAO,QAAQ;AAClB,cAAQ,QAAQ,eAAe,IAAI,UAAU,OAAO,MAAM;AAAA,IAC3D;AAEA,QAAI,aAAa,QAAQ;AACxB,aAAO,OAAO,QAAQ,IAAI,OAAO;AAAA,IAClC;AAEA,WAAO;AAAA,MACN,wCAAwC,OAAO,SAAS,IAAI,OAAO,SAAS;AAAA,MAC5E;AAAA,QACC,GAAG;AAAA,QACH,SAAS;AAAA,UACR,gBAAgB;AAAA,UAChB,GAAG;AAAA,UACH,GAAG;AAAA,UACH,GAAI,OAAO,WACR,EAAE,wBAAwB,UAAU,OAAO,QAAQ,GAAG,IACtD,CAAC;AAAA,QACL;AAAA,QACA,MAAM,KAAK,UAAU,OAAO;AAAA,MAC7B;AAAA,IACD;AAAA,EACD;AACD;AAgBA,SAAS,4BACR,UAC4B;AAC5B,SAAO,SAAS,IAAI,CAAC,QAAQ;AAC5B,UAAM,aAAa,EAAE,GAAG,IAAI;AAG5B,QAAI,WAAW,YAAY,QAAQ,WAAW,YAAY,QAAW;AACpE,iBAAW,UAAU;AAAA,IACtB;AAGA,QAAI,WAAW,gBAAgB,OAAO,WAAW,iBAAiB,UAAU;AAC3E,iBAAW,eAAe,mBAAmB,WAAW,YAAY;AAAA,IACrE;AAGA,QAAI,MAAM,QAAQ,WAAW,UAAU,GAAG;AACzC,iBAAW,aAAc,WAAW,WAAyC;AAAA,QAC5E,CAAC,OAAO;AACP,cAAI,GAAG,MAAM,OAAO,GAAG,OAAO,UAAU;AACvC,mBAAO,EAAE,GAAG,IAAI,IAAI,mBAAmB,GAAG,EAAE,EAAE;AAAA,UAC/C;AACA,iBAAO;AAAA,QACR;AAAA,MACD;AAAA,IACD;AAEA,WAAO;AAAA,EACR,CAAC;AACF;AAcA,SAAS,mBAAmB,IAAoB;AAC/C,QAAM,eAAe,GAAG,QAAQ,iBAAiB,EAAE;AAEnD,SAAO,aAAa,MAAM,GAAG,CAAC,EAAE,OAAO,GAAG,GAAG;AAC9C;AAWO,SAAS,4BAA4B,SAAyC;AACpF,SAAO,OAAO,QAAQ,SAAS;AAC9B,QAAI,CAAC,MAAM,MAAM;AAChB,aAAO,IAAI,SAAS,WAAW,EAAE,QAAQ,IAAI,CAAC;AAAA,IAC/C;AAEA,QAAI;AACJ,QAAI;AACH,aAAO,KAAK,MAAM,KAAK,IAAc;AAAA,IACtC,QAAQ;AACP,aAAO,IAAI,SAAS,qBAAqB,EAAE,QAAQ,IAAI,CAAC;AAAA,IACzD;AAEA,UAAM,QAAQ,KAAK;AACnB,UAAM,SAAS,KAAK;AAGpB,UAAM,SAAkC,CAAC;AACzC,QAAI,KAAK,UAAU;AAClB,aAAO,WAAW;AAAA,QACjB,KAAK;AAAA,MACN;AAAA,IACD;AACA,QAAI,KAAK,MAAO,QAAO,QAAQ,KAAK;AACpC,QAAI,OAAO,KAAK,gBAAgB,SAAU,QAAO,cAAc,KAAK;AACpE,QAAI,OAAO,KAAK,eAAe,SAAU,QAAO,aAAa,KAAK;AAClE,QAAI,KAAK,gBAAiB,QAAO,kBAAkB,KAAK;AACxD,QAAI,OAAQ,QAAO,SAAS;AAE5B,UAAM,SAAS,MAAM,QAAQ,IAAI,OAAO,MAAM;AAE9C,QAAI,UAAU,kBAAkB,gBAAgB;AAG/C,YAAM,cAAc;AAAA,QACnB;AAAA,QACA;AAAA,MACD;AACA,aAAO,IAAI,SAAS,aAAa;AAAA,QAChC,SAAS;AAAA,UACR,gBAAgB;AAAA,UAChB,iBAAiB;AAAA,QAClB;AAAA,MACD,CAAC;AAAA,IACF;AASA,UAAM,cACL,OAAO,WAAW,YAAY,WAAW,OACrC,SACD,EAAE,UAAU,OAAO,MAAM,EAAE;AAE/B,UAAM,eAAe,OAAO,YAAY,aAAa,WAAW,YAAY,WAAW;AAEvF,UAAM,UAAmC;AAAA,MACxC,MAAM;AAAA,MACN,SAAS;AAAA,IACV;AACA,QAAI,eAAe;AAGnB,QAAI,MAAM,QAAQ,YAAY,UAAU,KAAK,YAAY,WAAW,SAAS,GAAG;AAC/E,qBAAe;AACf,cAAQ,aAAa,YAAY,WAAW;AAAA,QAC3C,CAAC,QAKM;AAAA,UACN,IAAI,mBAAmB,GAAG,MAAM,OAAO,WAAW,CAAC;AAAA,UACnD,MAAM;AAAA,UACN,UAAU;AAAA,YACT,MAAM,GAAG,UAAU,QAAQ,GAAG,QAAQ;AAAA,YACtC,WACC,QAAQ,GAAG,UAAU,aAAa,GAAG,eAAe,WAC/C,GAAG,UAAU,aAAa,GAAG,YAC/B,KAAK,UAAU,GAAG,UAAU,aAAa,GAAG,aAAa,CAAC,CAAC;AAAA,UAChE;AAAA,QACD;AAAA,MACD;AAAA,IACD;AAEA,UAAM,iBAAiB;AAAA,MACtB,IAAI,cAAc,OAAO,WAAW,CAAC;AAAA,MACrC,QAAQ;AAAA,MACR,SAAS,KAAK,MAAM,KAAK,IAAI,IAAI,GAAI;AAAA,MACrC;AAAA,MACA,SAAS,CAAC,EAAE,OAAO,GAAG,SAAS,eAAe,aAAa,CAAC;AAAA,IAC7D;AAEA,WAAO,IAAI,SAAS,KAAK,UAAU,cAAc,GAAG;AAAA,MACnD,SAAS,EAAE,gBAAgB,mBAAmB;AAAA,IAC/C,CAAC;AAAA,EACF;AACD;AAgBA,SAAS,yBACR,QACA,OAC6B;AAC7B,QAAM,UAAU,IAAI,YAAY;AAChC,QAAM,UAAU,IAAI,YAAY;AAGhC,QAAM,WAAW,cAAc,OAAO,WAAW,CAAC;AAClD,QAAM,UAAU,KAAK,MAAM,KAAK,IAAI,IAAI,GAAI;AAC5C,MAAI,SAAS;AACb,MAAI,eAAe;AAInB,MAAI,iBAAiB;AAGrB,QAAM,uBAAuB,oBAAI,IAAY;AAE7C,SAAO,OAAO;AAAA,IACb,IAAI,gBAAwC;AAAA,MAC3C,UAAU,OAAO,YAAY;AAC5B,kBAAU,QAAQ,OAAO,OAAO,EAAE,QAAQ,KAAK,CAAC;AAChD,cAAM,QAAQ,OAAO,MAAM,IAAI;AAC/B,iBAAS,MAAM,IAAI,KAAK;AAExB,mBAAW,QAAQ,OAAO;AACzB,gBAAM,UAAU,KAAK,KAAK;AAC1B,cAAI,CAAC,WAAW,CAAC,QAAQ,WAAW,QAAQ,EAAG;AAC/C,gBAAM,OAAO,QAAQ,MAAM,CAAC;AAG5B,cAAI,SAAS,SAAU;AAEvB,cAAI;AACH,kBAAM,SAAS,KAAK,MAAM,IAAI;AAK9B,gBAAI,OAAO,YAAY,QAAW;AAGjC,+BAAiB;AACjB,oBAAM,SAAS,OAAO,UAAU,CAAC;AACjC,kBAAI,QAAQ,OAAO,YAAY;AAC9B,+BAAe;AACf,2BAAW,MAAM,OAAO,MAAM,YAAY;AACzC,sBAAI,GAAG,MAAM,OAAO,GAAG,OAAO,UAAU;AACvC,uBAAG,KAAK,mBAAmB,GAAG,EAAE;AAAA,kBACjC;AAAA,gBACD;AAAA,cACD;AACA,kBAAI,QAAQ,kBAAkB,cAAc;AAC3C,+BAAe;AAAA,cAChB;AACA,yBAAW;AAAA,gBACV,QAAQ,OAAO,SAAS,KAAK,UAAU,MAAM,CAAC;AAAA;AAAA,CAAM;AAAA,cACrD;AACA;AAAA,YACD;AAKA,gBAAI,OAAO,YAAY,QAAQ,OAAO,aAAa,IAAI;AACtD,oBAAM,cAAc;AAAA,gBACnB,IAAI;AAAA,gBACJ,QAAQ;AAAA,gBACR;AAAA,gBACA;AAAA,gBACA,SAAS;AAAA,kBACR;AAAA,oBACC,OAAO;AAAA,oBACP,OAAO,EAAE,SAAS,OAAO,SAAS;AAAA,oBAClC,eAAe;AAAA,kBAChB;AAAA,gBACD;AAAA,cACD;AACA,yBAAW;AAAA,gBACV,QAAQ,OAAO,SAAS,KAAK,UAAU,WAAW,CAAC;AAAA;AAAA,CAAM;AAAA,cAC1D;AAAA,YACD;AAOA,gBAAI,MAAM,QAAQ,OAAO,UAAU,KAAK,OAAO,WAAW,SAAS,GAAG;AACrE,yBAAW,MAAM,OAAO,YAAY;AACnC,sBAAM,UAAU,GAAG,SAAS;AAG5B,sBAAM,SAAS,GAAG,UAAU,QAAQ,GAAG,QAAQ;AAC/C,sBAAM,SAAS,GAAG,UAAU,aAAa,GAAG,aAAa;AACzD,sBAAM,OAAO,GAAG,MAAM;AAGtB,oBAAI,CAAC,QAAQ,CAAC,WAAW,CAAC,UAAU,WAAW,IAAK;AAEpD,+BAAe;AAGf,sBAAM,gBAAyC;AAAA,kBAC9C,OAAO;AAAA,gBACR;AAEA,oBAAI,CAAC,qBAAqB,IAAI,OAAO,GAAG;AAIvC,uCAAqB,IAAI,OAAO;AAChC,wBAAM,QAAQ,QAAQ,OAAO,QAAQ,GAAG,OAAO;AAC/C,gCAAc,KAAK,mBAAmB,KAAK;AAC3C,gCAAc,OAAO;AACrB,gCAAc,WAAW;AAAA,oBACxB,MAAM,UAAU;AAAA;AAAA,oBAEhB,WACC,UAAU,OACP,OAAO,WAAW,WACjB,SACA,KAAK,UAAU,MAAM,IACtB;AAAA,kBACL;AAAA,gBACD,OAAO;AAEN,sBAAI,UAAU,QAAQ,WAAW,IAAI;AACpC,kCAAc,WAAW;AAAA,sBACxB,WACC,OAAO,WAAW,WACf,SACA,KAAK,UAAU,MAAM;AAAA,oBAC1B;AAAA,kBACD,OAAO;AACN;AAAA,kBACD;AAAA,gBACD;AAEA,sBAAM,YAAY;AAAA,kBACjB,IAAI;AAAA,kBACJ,QAAQ;AAAA,kBACR;AAAA,kBACA;AAAA,kBACA,SAAS;AAAA,oBACR;AAAA,sBACC,OAAO;AAAA,sBACP,OAAO,EAAE,YAAY,CAAC,aAAa,EAAE;AAAA,sBACrC,eAAe;AAAA,oBAChB;AAAA,kBACD;AAAA,gBACD;AACA,2BAAW;AAAA,kBACV,QAAQ,OAAO,SAAS,KAAK,UAAU,SAAS,CAAC;AAAA;AAAA,CAAM;AAAA,gBACxD;AAAA,cACD;AAAA,YACD;AAAA,UACD,SAAS,GAAG;AAEX,oBAAQ,KAAK,4CAA4C,MAAM,CAAC;AAAA,UACjE;AAAA,QACD;AAAA,MACD;AAAA,MACA,MAAM,YAAY;AACjB,YAAI,CAAC,gBAAgB;AAEpB,gBAAM,aAAa;AAAA,YAClB,IAAI;AAAA,YACJ,QAAQ;AAAA,YACR;AAAA,YACA;AAAA,YACA,SAAS;AAAA,cACR;AAAA,gBACC,OAAO;AAAA,gBACP,OAAO,CAAC;AAAA,gBACR,eAAe,eAAe,eAAe;AAAA,cAC9C;AAAA,YACD;AAAA,UACD;AACA,qBAAW,QAAQ,QAAQ,OAAO,SAAS,KAAK,UAAU,UAAU,CAAC;AAAA;AAAA,CAAM,CAAC;AAAA,QAC7E;AAGA,mBAAW,QAAQ,QAAQ,OAAO,kBAAkB,CAAC;AAAA,MACtD;AAAA,IACD,CAAC;AAAA,EACF;AACD;","names":[]}

package/dist/chunk-V6TY7KAL.js

@@ -0,0 +1,8 @@
+var __defProp = Object.defineProperty;
+var __defNormalProp = (obj, key, value) => key in obj ? __defProp(obj, key, { enumerable: true, configurable: true, writable: true, value }) : obj[key] = value;
+var __publicField = (obj, key, value) => __defNormalProp(obj, typeof key !== "symbol" ? key + "" : key, value);
+
+export {
+  __publicField
+};
+//# sourceMappingURL=chunk-V6TY7KAL.js.map

package/dist/chunk-V6TY7KAL.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":[],"sourcesContent":[],"mappings":"","names":[]}

package/dist/chunk-VV3JFKAN.js

@@ -0,0 +1,216 @@
+import {
+  uint8ArrayToBase64
+} from "./chunk-XU7YEPML.js";
+import {
+  workersAiRestFetch,
+  workersAiRestFetchBinary
+} from "./chunk-2VII5BK2.js";
+import {
+  createGatewayFetch,
+  isDirectBindingConfig,
+  isDirectCredentialsConfig
+} from "./chunk-UUFEOQ6B.js";
+import {
+  __publicField
+} from "./chunk-V6TY7KAL.js";
+
+// src/adapters/workers-ai-transcription.ts
+import { BaseTranscriptionAdapter } from "@tanstack/ai/adapters";
+var WorkersAiTranscriptionAdapter = class extends BaseTranscriptionAdapter {
+  constructor(config, model) {
+    super({}, model);
+    __publicField(this, "name", "workers-ai-transcription");
+    __publicField(this, "adapterConfig");
+    this.adapterConfig = config;
+  }
+  async transcribe(options) {
+    const { audio, language, prompt, modelOptions } = options;
+    const audioBytes = await normalizeAudioToBytes(audio);
+    const extra = { ...modelOptions };
+    if (language) extra.language = language;
+    if (prompt) extra.initial_prompt = prompt;
+    const audioPayload = this.buildAudioPayload(audioBytes, audio);
+    if (isDirectBindingConfig(this.adapterConfig)) {
+      return this.transcribeViaBinding(audioPayload, extra);
+    }
+    if (isDirectCredentialsConfig(this.adapterConfig)) {
+      if (this.model === "@cf/deepgram/nova-3") {
+        return this.transcribeViaRestBinary(audioBytes, audio, extra);
+      }
+      return this.transcribeViaRest(audioPayload, extra);
+    }
+    return this.transcribeViaGateway(audioPayload, extra);
+  }
+  /**
+   * Build the audio field for the request payload, handling model-specific formats.
+   *
+   * - `@cf/deepgram/nova-3` requires `{ body: base64, contentType: "audio/..." }`
+   * - `@cf/openai/whisper-large-v3-turbo` REST/gateway accepts a base64 string
+   * - Other Whisper models accept `number[]` (binding) or base64 (REST)
+   */
+  buildAudioPayload(audioBytes, originalAudio) {
+    if (this.model === "@cf/deepgram/nova-3") {
+      const b64 = uint8ArrayToBase64(new Uint8Array(audioBytes));
+      const contentType = detectAudioContentType(originalAudio);
+      return { audio: { body: b64, contentType } };
+    }
+    if (this.model === "@cf/openai/whisper-large-v3-turbo") {
+      return { audio: uint8ArrayToBase64(new Uint8Array(audioBytes)) };
+    }
+    return { audio: audioBytes };
+  }
+  async transcribeViaBinding(audioPayload, options) {
+    const ai = this.adapterConfig.binding;
+    const result = await ai.run(this.model, {
+      ...audioPayload,
+      ...options
+    });
+    return this.normalizeResult(result);
+  }
+  async transcribeViaRest(audioPayload, options) {
+    const config = this.adapterConfig;
+    const response = await workersAiRestFetch(
+      config,
+      this.model,
+      { ...audioPayload, ...options },
+      {
+        label: "Workers AI transcription",
+        signal: options.signal
+      }
+    );
+    const data = await response.json();
+    return this.normalizeResult(data.result ?? data);
+  }
+  /**
+   * Transcribe via REST using raw binary audio.
+   * Required for models like Deepgram Nova-3 that expect raw audio bytes
+   * with a Content-Type header (e.g. "audio/wav") instead of JSON.
+   */
+  async transcribeViaRestBinary(audioBytes, originalAudio, options) {
+    const config = this.adapterConfig;
+    const contentType = detectAudioContentType(originalAudio);
+    const response = await workersAiRestFetchBinary(
+      config,
+      this.model,
+      new Uint8Array(audioBytes),
+      contentType,
+      {
+        label: "Workers AI transcription",
+        signal: options.signal
+      }
+    );
+    const data = await response.json();
+    return this.normalizeResult(data.result ?? data);
+  }
+  async transcribeViaGateway(audioPayload, options) {
+    const gatewayConfig = this.adapterConfig;
+    const gatewayFetch = createGatewayFetch("workers-ai", gatewayConfig);
+    const response = await gatewayFetch("https://api.cloudflare.com/v1/audio/transcriptions", {
+      method: "POST",
+      body: JSON.stringify({
+        model: this.model,
+        ...audioPayload,
+        ...options
+      })
+    });
+    if (!response.ok) {
+      const errorText = await response.text();
+      throw new Error(
+        `Workers AI transcription gateway request failed (${response.status}): ${errorText}`
+      );
+    }
+    const data = await response.json();
+    return this.normalizeResult(data);
+  }
+  /**
+   * Normalize Workers AI transcription results into the standard
+   * TanStack AI TranscriptionResult shape.
+   *
+   * Handles three response formats:
+   * - Whisper: `{ text, words?, vtt? }`
+   * - Whisper v3-turbo: `{ text, segments?, transcription_info? }`
+   * - Deepgram Nova-3: `{ results: { channels: [{ alternatives: [{ transcript, words }] }] } }`
+   */
+  normalizeResult(raw) {
+    const results = raw.results;
+    if (results?.channels) {
+      const channels = results.channels;
+      const alt = channels?.[0]?.alternatives?.[0];
+      const text = alt?.transcript ?? "";
+      const result2 = {
+        id: this.generateId(),
+        model: this.model,
+        text
+      };
+      if (alt?.words && Array.isArray(alt.words)) {
+        result2.words = alt.words.map((w) => ({
+          word: w.word ?? "",
+          start: w.start ?? 0,
+          end: w.end ?? 0
+        }));
+      }
+      return result2;
+    }
+    const result = {
+      id: this.generateId(),
+      model: this.model,
+      text: raw.text ?? ""
+    };
+    const transcriptionInfo = raw.transcription_info;
+    if (transcriptionInfo?.language) {
+      result.language = transcriptionInfo.language;
+    }
+    if (transcriptionInfo?.duration != null) {
+      result.duration = transcriptionInfo.duration;
+    }
+    if (raw.segments && Array.isArray(raw.segments)) {
+      result.segments = raw.segments.map((seg, idx) => ({
+        id: idx,
+        text: seg.text ?? "",
+        start: seg.start ?? 0,
+        end: seg.end ?? 0
+      }));
+    }
+    if (raw.words && Array.isArray(raw.words)) {
+      result.words = raw.words.map((w) => ({
+        word: w.word ?? "",
+        start: w.start ?? 0,
+        end: w.end ?? 0
+      }));
+    }
+    return result;
+  }
+};
+function createWorkersAiTranscription(model, config) {
+  return new WorkersAiTranscriptionAdapter(config, model);
+}
+async function normalizeAudioToBytes(audio) {
+  if (audio instanceof ArrayBuffer) {
+    return Array.from(new Uint8Array(audio));
+  }
+  if (audio instanceof Blob) {
+    const buffer = await audio.arrayBuffer();
+    return Array.from(new Uint8Array(buffer));
+  }
+  if (typeof audio === "string") {
+    const binary = atob(audio);
+    const bytes = new Uint8Array(binary.length);
+    for (let i = 0; i < binary.length; i++) {
+      bytes[i] = binary.charCodeAt(i);
+    }
+    return Array.from(bytes);
+  }
+  throw new Error("Unsupported audio format. Expected string, File, Blob, or ArrayBuffer.");
+}
+function detectAudioContentType(audio) {
+  if (audio instanceof Blob && audio.type) {
+    return audio.type;
+  }
+  return "audio/wav";
+}
+
+export {
+  WorkersAiTranscriptionAdapter,
+  createWorkersAiTranscription
+};
+//# sourceMappingURL=chunk-VV3JFKAN.js.map

package/dist/chunk-VV3JFKAN.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/adapters/workers-ai-transcription.ts"],"sourcesContent":["import { BaseTranscriptionAdapter } from \"@tanstack/ai/adapters\";\nimport type { TranscriptionOptions, TranscriptionResult } from \"@tanstack/ai\";\nimport {\n\ttype WorkersAiAdapterConfig,\n\ttype WorkersAiDirectBindingConfig,\n\ttype WorkersAiDirectCredentialsConfig,\n\ttype AiGatewayAdapterConfig,\n\tcreateGatewayFetch,\n\tisDirectBindingConfig,\n\tisDirectCredentialsConfig,\n} from \"../utils/create-fetcher\";\nimport { workersAiRestFetch, workersAiRestFetchBinary } from \"../utils/workers-ai-rest\";\nimport { uint8ArrayToBase64 } from \"../utils/binary\";\n\n// ---------------------------------------------------------------------------\n// Model types\n// ---------------------------------------------------------------------------\n\n/**\n * Workers AI models that support speech-to-text transcription.\n *\n * Note: the typed `AiModels` interface in `@cloudflare/workers-types` may lag\n * behind what's deployed. We use a string union here that matches the known\n * models including Deepgram partner models.\n *\n * **Nova-3 note:** `@cf/deepgram/nova-3` uses a different input format than the\n * Whisper models. Via binding it accepts `{ audio: { body: base64, contentType } }`.\n * Via REST it requires multipart form data (not JSON). The adapter handles both.\n */\nexport type WorkersAiTranscriptionModel =\n\t| \"@cf/openai/whisper\"\n\t| \"@cf/openai/whisper-tiny-en\"\n\t| \"@cf/openai/whisper-large-v3-turbo\"\n\t| \"@cf/deepgram/nova-3\";\n\n// ---------------------------------------------------------------------------\n// WorkersAiTranscriptionAdapter\n// ---------------------------------------------------------------------------\n\nexport class WorkersAiTranscriptionAdapter extends BaseTranscriptionAdapter<WorkersAiTranscriptionModel> {\n\treadonly name = \"workers-ai-transcription\" as const;\n\tprivate adapterConfig: WorkersAiAdapterConfig;\n\n\tconstructor(config: WorkersAiAdapterConfig, model: WorkersAiTranscriptionModel) {\n\t\tsuper({}, model);\n\t\tthis.adapterConfig = config;\n\t}\n\n\tasync transcribe(options: TranscriptionOptions): Promise<TranscriptionResult> {\n\t\tconst { audio, language, prompt, modelOptions } = options;\n\n\t\t// Normalize audio to raw bytes\n\t\tconst audioBytes = await normalizeAudioToBytes(audio);\n\n\t\tconst extra: Record<string, unknown> = { ...modelOptions };\n\t\tif (language) extra.language = language;\n\t\tif (prompt) extra.initial_prompt = prompt;\n\n\t\t// Build the model-specific audio payload:\n\t\t// - Deepgram Nova-3 (binding): { audio: { body: base64, contentType: \"audio/...\" } }\n\t\t// - Deepgram Nova-3 (REST): multipart FormData (handled separately)\n\t\t// - Whisper Large v3 Turbo (REST/gateway): { audio: base64string }\n\t\t// - Other Whisper models (binding): { audio: number[] }\n\t\tconst audioPayload = this.buildAudioPayload(audioBytes, audio);\n\n\t\tif (isDirectBindingConfig(this.adapterConfig)) {\n\t\t\treturn this.transcribeViaBinding(audioPayload, extra);\n\t\t}\n\n\t\tif (isDirectCredentialsConfig(this.adapterConfig)) {\n\t\t\t// Nova-3 REST requires raw binary audio, not JSON\n\t\t\tif (this.model === \"@cf/deepgram/nova-3\") {\n\t\t\t\treturn this.transcribeViaRestBinary(audioBytes, audio, extra);\n\t\t\t}\n\t\t\treturn this.transcribeViaRest(audioPayload, extra);\n\t\t}\n\n\t\treturn this.transcribeViaGateway(audioPayload, extra);\n\t}\n\n\t/**\n\t * Build the audio field for the request payload, handling model-specific formats.\n\t *\n\t * - `@cf/deepgram/nova-3` requires `{ body: base64, contentType: \"audio/...\" }`\n\t * - `@cf/openai/whisper-large-v3-turbo` REST/gateway accepts a base64 string\n\t * - Other Whisper models accept `number[]` (binding) or base64 (REST)\n\t */\n\tprivate buildAudioPayload(\n\t\taudioBytes: number[],\n\t\toriginalAudio: string | File | Blob | ArrayBuffer,\n\t): Record<string, unknown> {\n\t\tif (this.model === \"@cf/deepgram/nova-3\") {\n\t\t\tconst b64 = uint8ArrayToBase64(new Uint8Array(audioBytes));\n\t\t\tconst contentType = detectAudioContentType(originalAudio);\n\t\t\treturn { audio: { body: b64, contentType } };\n\t\t}\n\n\t\tif (this.model === \"@cf/openai/whisper-large-v3-turbo\") {\n\t\t\treturn { audio: uint8ArrayToBase64(new Uint8Array(audioBytes)) };\n\t\t}\n\n\t\treturn { audio: audioBytes };\n\t}\n\n\tprivate async transcribeViaBinding(\n\t\taudioPayload: Record<string, unknown>,\n\t\toptions: Record<string, unknown>,\n\t): Promise<TranscriptionResult> {\n\t\tconst ai = (this.adapterConfig as WorkersAiDirectBindingConfig).binding;\n\t\tconst result = (await ai.run(this.model, {\n\t\t\t...audioPayload,\n\t\t\t...options,\n\t\t})) as Record<string, unknown>;\n\t\treturn this.normalizeResult(result);\n\t}\n\n\tprivate async transcribeViaRest(\n\t\taudioPayload: Record<string, unknown>,\n\t\toptions: Record<string, unknown>,\n\t): Promise<TranscriptionResult> {\n\t\tconst config = this.adapterConfig as WorkersAiDirectCredentialsConfig;\n\n\t\tconst response = await workersAiRestFetch(\n\t\t\tconfig,\n\t\t\tthis.model,\n\t\t\t{ ...audioPayload, ...options },\n\t\t\t{\n\t\t\t\tlabel: \"Workers AI transcription\",\n\t\t\t\tsignal: (options as { signal?: AbortSignal }).signal,\n\t\t\t},\n\t\t);\n\n\t\tconst data = (await response.json()) as {\n\t\t\tresult?: Record<string, unknown>;\n\t\t} & Record<string, unknown>;\n\n\t\t// Cloudflare REST API wraps responses in { success, result: {...} }.\n\t\t// Use `data.result` when present, fall back to `data` for direct responses.\n\t\treturn this.normalizeResult(data.result ?? data);\n\t}\n\n\t/**\n\t * Transcribe via REST using raw binary audio.\n\t * Required for models like Deepgram Nova-3 that expect raw audio bytes\n\t * with a Content-Type header (e.g. \"audio/wav\") instead of JSON.\n\t */\n\tprivate async transcribeViaRestBinary(\n\t\taudioBytes: number[],\n\t\toriginalAudio: string | File | Blob | ArrayBuffer,\n\t\toptions: Record<string, unknown>,\n\t): Promise<TranscriptionResult> {\n\t\tconst config = this.adapterConfig as WorkersAiDirectCredentialsConfig;\n\t\tconst contentType = detectAudioContentType(originalAudio);\n\n\t\tconst response = await workersAiRestFetchBinary(\n\t\t\tconfig,\n\t\t\tthis.model,\n\t\t\tnew Uint8Array(audioBytes),\n\t\t\tcontentType,\n\t\t\t{\n\t\t\t\tlabel: \"Workers AI transcription\",\n\t\t\t\tsignal: (options as { signal?: AbortSignal }).signal,\n\t\t\t},\n\t\t);\n\n\t\tconst data = (await response.json()) as {\n\t\t\tresult?: Record<string, unknown>;\n\t\t} & Record<string, unknown>;\n\n\t\treturn this.normalizeResult(data.result ?? data);\n\t}\n\n\tprivate async transcribeViaGateway(\n\t\taudioPayload: Record<string, unknown>,\n\t\toptions: Record<string, unknown>,\n\t): Promise<TranscriptionResult> {\n\t\tconst gatewayConfig = this.adapterConfig as AiGatewayAdapterConfig;\n\t\tconst gatewayFetch = createGatewayFetch(\"workers-ai\", gatewayConfig);\n\n\t\t// The URL here is a placeholder — createGatewayFetch for \"workers-ai\" extracts\n\t\t// the model from the body, sets it as the endpoint, and routes through the gateway.\n\t\t// The actual URL path is not used.\n\t\tconst response = await gatewayFetch(\"https://api.cloudflare.com/v1/audio/transcriptions\", {\n\t\t\tmethod: \"POST\",\n\t\t\tbody: JSON.stringify({\n\t\t\t\tmodel: this.model,\n\t\t\t\t...audioPayload,\n\t\t\t\t...options,\n\t\t\t}),\n\t\t});\n\n\t\tif (!response.ok) {\n\t\t\tconst errorText = await response.text();\n\t\t\tthrow new Error(\n\t\t\t\t`Workers AI transcription gateway request failed (${response.status}): ${errorText}`,\n\t\t\t);\n\t\t}\n\n\t\tconst data = (await response.json()) as Record<string, unknown>;\n\t\treturn this.normalizeResult(data);\n\t}\n\n\t/**\n\t * Normalize Workers AI transcription results into the standard\n\t * TanStack AI TranscriptionResult shape.\n\t *\n\t * Handles three response formats:\n\t * - Whisper: `{ text, words?, vtt? }`\n\t * - Whisper v3-turbo: `{ text, segments?, transcription_info? }`\n\t * - Deepgram Nova-3: `{ results: { channels: [{ alternatives: [{ transcript, words }] }] } }`\n\t */\n\tprivate normalizeResult(raw: Record<string, unknown>): TranscriptionResult {\n\t\t// Deepgram Nova-3 format: { results: { channels: [{ alternatives: [{ transcript, words }] }] } }\n\t\tconst results = raw.results as Record<string, unknown> | undefined;\n\t\tif (results?.channels) {\n\t\t\tconst channels = results.channels as Array<{\n\t\t\t\talternatives?: Array<{\n\t\t\t\t\ttranscript?: string;\n\t\t\t\t\tconfidence?: number;\n\t\t\t\t\twords?: Array<{ word: string; start: number; end: number; confidence: number }>;\n\t\t\t\t}>;\n\t\t\t}>;\n\t\t\tconst alt = channels?.[0]?.alternatives?.[0];\n\t\t\tconst text = alt?.transcript ?? \"\";\n\t\t\tconst result: TranscriptionResult = {\n\t\t\t\tid: this.generateId(),\n\t\t\t\tmodel: this.model,\n\t\t\t\ttext,\n\t\t\t};\n\t\t\tif (alt?.words && Array.isArray(alt.words)) {\n\t\t\t\tresult.words = alt.words.map((w) => ({\n\t\t\t\t\tword: w.word ?? \"\",\n\t\t\t\t\tstart: w.start ?? 0,\n\t\t\t\t\tend: w.end ?? 0,\n\t\t\t\t}));\n\t\t\t}\n\t\t\treturn result;\n\t\t}\n\n\t\t// Whisper format: { text, words?, vtt? }\n\t\t// Whisper v3-turbo format: { text, segments?, transcription_info? }\n\t\tconst result: TranscriptionResult = {\n\t\t\tid: this.generateId(),\n\t\t\tmodel: this.model,\n\t\t\ttext: (raw.text as string) ?? \"\",\n\t\t};\n\n\t\t// Language from transcription_info (whisper-large-v3-turbo)\n\t\tconst transcriptionInfo = raw.transcription_info as Record<string, unknown> | undefined;\n\t\tif (transcriptionInfo?.language) {\n\t\t\tresult.language = transcriptionInfo.language as string;\n\t\t}\n\n\t\t// Duration\n\t\tif (transcriptionInfo?.duration != null) {\n\t\t\tresult.duration = transcriptionInfo.duration as number;\n\t\t}\n\n\t\t// Segments (whisper-large-v3-turbo returns these)\n\t\tif (raw.segments && Array.isArray(raw.segments)) {\n\t\t\tresult.segments = raw.segments.map((seg: Record<string, unknown>, idx: number) => ({\n\t\t\t\tid: idx,\n\t\t\t\ttext: (seg.text as string) ?? \"\",\n\t\t\t\tstart: (seg.start as number) ?? 0,\n\t\t\t\tend: (seg.end as number) ?? 0,\n\t\t\t}));\n\t\t}\n\n\t\t// Words — basic whisper returns top-level words[], v3-turbo nests them in segments\n\t\tif (raw.words && Array.isArray(raw.words)) {\n\t\t\tresult.words = raw.words.map((w: Record<string, unknown>) => ({\n\t\t\t\tword: (w.word as string) ?? \"\",\n\t\t\t\tstart: (w.start as number) ?? 0,\n\t\t\t\tend: (w.end as number) ?? 0,\n\t\t\t}));\n\t\t}\n\n\t\treturn result;\n\t}\n}\n\n// ---------------------------------------------------------------------------\n// Factory function\n// ---------------------------------------------------------------------------\n\n/**\n * Creates a Workers AI transcription adapter for speech-to-text.\n *\n * Works with TanStack AI's `generateTranscription()` activity function:\n * ```ts\n * import { generateTranscription } from \"@tanstack/ai\";\n * import { createWorkersAiTranscription } from \"@cloudflare/tanstack-ai\";\n *\n * const adapter = createWorkersAiTranscription(\n *   \"@cf/openai/whisper-large-v3-turbo\",\n *   { binding: env.AI },\n * );\n *\n * const result = await generateTranscription({ adapter, audio: audioData });\n * // result.text — the transcribed text\n * ```\n *\n * Note: Factory takes `(model, config)` for ergonomics — the class constructor\n * uses `(config, model)` to match TanStack AI's upstream convention.\n */\nexport function createWorkersAiTranscription(\n\tmodel: WorkersAiTranscriptionModel,\n\tconfig: WorkersAiAdapterConfig,\n) {\n\treturn new WorkersAiTranscriptionAdapter(config, model);\n}\n\n// ---------------------------------------------------------------------------\n// Utilities\n// ---------------------------------------------------------------------------\n\n/**\n * Normalize various audio input formats into a number[] (raw bytes).\n *\n * Note: `File extends Blob`, so `File` instances are handled by the\n * `instanceof Blob` branch. `Blob.arrayBuffer()` always reads the full\n * contents regardless of any prior reads — there's no cursor to worry about.\n */\nasync function normalizeAudioToBytes(audio: string | File | Blob | ArrayBuffer): Promise<number[]> {\n\tif (audio instanceof ArrayBuffer) {\n\t\treturn Array.from(new Uint8Array(audio));\n\t}\n\n\tif (audio instanceof Blob) {\n\t\t// This also handles `File` (which extends Blob)\n\t\tconst buffer = await audio.arrayBuffer();\n\t\treturn Array.from(new Uint8Array(buffer));\n\t}\n\n\tif (typeof audio === \"string\") {\n\t\t// Assume base64 string — decode to bytes\n\t\tconst binary = atob(audio);\n\t\tconst bytes = new Uint8Array(binary.length);\n\t\tfor (let i = 0; i < binary.length; i++) {\n\t\t\tbytes[i] = binary.charCodeAt(i);\n\t\t}\n\t\treturn Array.from(bytes);\n\t}\n\n\tthrow new Error(\"Unsupported audio format. Expected string, File, Blob, or ArrayBuffer.\");\n}\n\n/**\n * Detect the MIME type of the audio input for models that require it\n * (e.g., Deepgram Nova-3).\n *\n * - `File` / `Blob`: use the `.type` property (e.g., \"audio/wav\")\n * - `ArrayBuffer` / `string`: defaults to \"audio/wav\"\n */\nfunction detectAudioContentType(audio: string | File | Blob | ArrayBuffer): string {\n\t// File and Blob carry their own MIME type\n\tif (audio instanceof Blob && audio.type) {\n\t\treturn audio.type;\n\t}\n\n\t// For raw bytes, default to audio/wav — this is the most common\n\t// format for transcription inputs and what the E2E tests use.\n\treturn \"audio/wav\";\n}\n"],"mappings":";;;;;;;;;;;;;;;;;AAAA,SAAS,gCAAgC;AAuClC,IAAM,gCAAN,cAA4C,yBAAsD;AAAA,EAIxG,YAAY,QAAgC,OAAoC;AAC/E,UAAM,CAAC,GAAG,KAAK;AAJhB,wBAAS,QAAO;AAChB,wBAAQ;AAIP,SAAK,gBAAgB;AAAA,EACtB;AAAA,EAEA,MAAM,WAAW,SAA6D;AAC7E,UAAM,EAAE,OAAO,UAAU,QAAQ,aAAa,IAAI;AAGlD,UAAM,aAAa,MAAM,sBAAsB,KAAK;AAEpD,UAAM,QAAiC,EAAE,GAAG,aAAa;AACzD,QAAI,SAAU,OAAM,WAAW;AAC/B,QAAI,OAAQ,OAAM,iBAAiB;AAOnC,UAAM,eAAe,KAAK,kBAAkB,YAAY,KAAK;AAE7D,QAAI,sBAAsB,KAAK,aAAa,GAAG;AAC9C,aAAO,KAAK,qBAAqB,cAAc,KAAK;AAAA,IACrD;AAEA,QAAI,0BAA0B,KAAK,aAAa,GAAG;AAElD,UAAI,KAAK,UAAU,uBAAuB;AACzC,eAAO,KAAK,wBAAwB,YAAY,OAAO,KAAK;AAAA,MAC7D;AACA,aAAO,KAAK,kBAAkB,cAAc,KAAK;AAAA,IAClD;AAEA,WAAO,KAAK,qBAAqB,cAAc,KAAK;AAAA,EACrD;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASQ,kBACP,YACA,eAC0B;AAC1B,QAAI,KAAK,UAAU,uBAAuB;AACzC,YAAM,MAAM,mBAAmB,IAAI,WAAW,UAAU,CAAC;AACzD,YAAM,cAAc,uBAAuB,aAAa;AACxD,aAAO,EAAE,OAAO,EAAE,MAAM,KAAK,YAAY,EAAE;AAAA,IAC5C;AAEA,QAAI,KAAK,UAAU,qCAAqC;AACvD,aAAO,EAAE,OAAO,mBAAmB,IAAI,WAAW,UAAU,CAAC,EAAE;AAAA,IAChE;AAEA,WAAO,EAAE,OAAO,WAAW;AAAA,EAC5B;AAAA,EAEA,MAAc,qBACb,cACA,SAC+B;AAC/B,UAAM,KAAM,KAAK,cAA+C;AAChE,UAAM,SAAU,MAAM,GAAG,IAAI,KAAK,OAAO;AAAA,MACxC,GAAG;AAAA,MACH,GAAG;AAAA,IACJ,CAAC;AACD,WAAO,KAAK,gBAAgB,MAAM;AAAA,EACnC;AAAA,EAEA,MAAc,kBACb,cACA,SAC+B;AAC/B,UAAM,SAAS,KAAK;AAEpB,UAAM,WAAW,MAAM;AAAA,MACtB;AAAA,MACA,KAAK;AAAA,MACL,EAAE,GAAG,cAAc,GAAG,QAAQ;AAAA,MAC9B;AAAA,QACC,OAAO;AAAA,QACP,QAAS,QAAqC;AAAA,MAC/C;AAAA,IACD;AAEA,UAAM,OAAQ,MAAM,SAAS,KAAK;AAMlC,WAAO,KAAK,gBAAgB,KAAK,UAAU,IAAI;AAAA,EAChD;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,MAAc,wBACb,YACA,eACA,SAC+B;AAC/B,UAAM,SAAS,KAAK;AACpB,UAAM,cAAc,uBAAuB,aAAa;AAExD,UAAM,WAAW,MAAM;AAAA,MACtB;AAAA,MACA,KAAK;AAAA,MACL,IAAI,WAAW,UAAU;AAAA,MACzB;AAAA,MACA;AAAA,QACC,OAAO;AAAA,QACP,QAAS,QAAqC;AAAA,MAC/C;AAAA,IACD;AAEA,UAAM,OAAQ,MAAM,SAAS,KAAK;AAIlC,WAAO,KAAK,gBAAgB,KAAK,UAAU,IAAI;AAAA,EAChD;AAAA,EAEA,MAAc,qBACb,cACA,SAC+B;AAC/B,UAAM,gBAAgB,KAAK;AAC3B,UAAM,eAAe,mBAAmB,cAAc,aAAa;AAKnE,UAAM,WAAW,MAAM,aAAa,sDAAsD;AAAA,MACzF,QAAQ;AAAA,MACR,MAAM,KAAK,UAAU;AAAA,QACpB,OAAO,KAAK;AAAA,QACZ,GAAG;AAAA,QACH,GAAG;AAAA,MACJ,CAAC;AAAA,IACF,CAAC;AAED,QAAI,CAAC,SAAS,IAAI;AACjB,YAAM,YAAY,MAAM,SAAS,KAAK;AACtC,YAAM,IAAI;AAAA,QACT,oDAAoD,SAAS,MAAM,MAAM,SAAS;AAAA,MACnF;AAAA,IACD;AAEA,UAAM,OAAQ,MAAM,SAAS,KAAK;AAClC,WAAO,KAAK,gBAAgB,IAAI;AAAA,EACjC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAWQ,gBAAgB,KAAmD;AAE1E,UAAM,UAAU,IAAI;AACpB,QAAI,SAAS,UAAU;AACtB,YAAM,WAAW,QAAQ;AAOzB,YAAM,MAAM,WAAW,CAAC,GAAG,eAAe,CAAC;AAC3C,YAAM,OAAO,KAAK,cAAc;AAChC,YAAMA,UAA8B;AAAA,QACnC,IAAI,KAAK,WAAW;AAAA,QACpB,OAAO,KAAK;AAAA,QACZ;AAAA,MACD;AACA,UAAI,KAAK,SAAS,MAAM,QAAQ,IAAI,KAAK,GAAG;AAC3C,QAAAA,QAAO,QAAQ,IAAI,MAAM,IAAI,CAAC,OAAO;AAAA,UACpC,MAAM,EAAE,QAAQ;AAAA,UAChB,OAAO,EAAE,SAAS;AAAA,UAClB,KAAK,EAAE,OAAO;AAAA,QACf,EAAE;AAAA,MACH;AACA,aAAOA;AAAA,IACR;AAIA,UAAM,SAA8B;AAAA,MACnC,IAAI,KAAK,WAAW;AAAA,MACpB,OAAO,KAAK;AAAA,MACZ,MAAO,IAAI,QAAmB;AAAA,IAC/B;AAGA,UAAM,oBAAoB,IAAI;AAC9B,QAAI,mBAAmB,UAAU;AAChC,aAAO,WAAW,kBAAkB;AAAA,IACrC;AAGA,QAAI,mBAAmB,YAAY,MAAM;AACxC,aAAO,WAAW,kBAAkB;AAAA,IACrC;AAGA,QAAI,IAAI,YAAY,MAAM,QAAQ,IAAI,QAAQ,GAAG;AAChD,aAAO,WAAW,IAAI,SAAS,IAAI,CAAC,KAA8B,SAAiB;AAAA,QAClF,IAAI;AAAA,QACJ,MAAO,IAAI,QAAmB;AAAA,QAC9B,OAAQ,IAAI,SAAoB;AAAA,QAChC,KAAM,IAAI,OAAkB;AAAA,MAC7B,EAAE;AAAA,IACH;AAGA,QAAI,IAAI,SAAS,MAAM,QAAQ,IAAI,KAAK,GAAG;AAC1C,aAAO,QAAQ,IAAI,MAAM,IAAI,CAAC,OAAgC;AAAA,QAC7D,MAAO,EAAE,QAAmB;AAAA,QAC5B,OAAQ,EAAE,SAAoB;AAAA,QAC9B,KAAM,EAAE,OAAkB;AAAA,MAC3B,EAAE;AAAA,IACH;AAEA,WAAO;AAAA,EACR;AACD;AA0BO,SAAS,6BACf,OACA,QACC;AACD,SAAO,IAAI,8BAA8B,QAAQ,KAAK;AACvD;AAaA,eAAe,sBAAsB,OAA8D;AAClG,MAAI,iBAAiB,aAAa;AACjC,WAAO,MAAM,KAAK,IAAI,WAAW,KAAK,CAAC;AAAA,EACxC;AAEA,MAAI,iBAAiB,MAAM;AAE1B,UAAM,SAAS,MAAM,MAAM,YAAY;AACvC,WAAO,MAAM,KAAK,IAAI,WAAW,MAAM,CAAC;AAAA,EACzC;AAEA,MAAI,OAAO,UAAU,UAAU;AAE9B,UAAM,SAAS,KAAK,KAAK;AACzB,UAAM,QAAQ,IAAI,WAAW,OAAO,MAAM;AAC1C,aAAS,IAAI,GAAG,IAAI,OAAO,QAAQ,KAAK;AACvC,YAAM,CAAC,IAAI,OAAO,WAAW,CAAC;AAAA,IAC/B;AACA,WAAO,MAAM,KAAK,KAAK;AAAA,EACxB;AAEA,QAAM,IAAI,MAAM,wEAAwE;AACzF;AASA,SAAS,uBAAuB,OAAmD;AAElF,MAAI,iBAAiB,QAAQ,MAAM,MAAM;AACxC,WAAO,MAAM;AAAA,EACd;AAIA,SAAO;AACR;","names":["result"]}