use authorization header instead of env token

src/routes/api/+server.ts

@@ -1,34 +1,31 @@
 import { json, type RequestEvent } from '@sveltejs/kit';
 import { InferenceClient } from '@huggingface/inference';
 
-import { env } from '$env/dynamic/private';
-
 export async function POST({ request }: RequestEvent) {
 	const { model, messages, options, provider = 'auto' } = await request.json();
+	const token = request.headers.get('Authorization')?.split(' ')[1];
+	if (!token) {
+		return json({ error: 'Unauthorized' }, { status: 401 });
+	}
 
 	if (!model) {
 		return json({ error: 'Model and prompt are required' }, { status: 400 });
 	}
 
-	const client = new InferenceClient(env.DEFAULT_HF_TOKEN);
+	const client = new InferenceClient(token);
 
-	const stream = client.chatCompletionStream(
-		{
-			model: model + (provider !== 'auto' ? `:${provider}` : ''),
-			...(options ?? {}),
-			messages: [
-				{
-					role: 'system',
-					content:
-						"You are a helpful assistant. You are very helpful and friendly. Use markdown to format your responses, but don't include array start and end markers."
-				},
-				...(messages ?? [])
-			]
-		},
-		{
-			billTo: 'huggingface'
-		}
-	);
+	const stream = client.chatCompletionStream({
+		model: model + (provider !== 'auto' ? `:${provider}` : ''),
+		...(options ?? {}),
+		messages: [
+			{
+				role: 'system',
+				content:
+					"You are a helpful assistant. You are very helpful and friendly. Use markdown to format your responses, but don't include array start and end markers."
+			},
+			...(messages ?? [])
+		]
+	});
 
 	const readable = new ReadableStream({
 		async start(controller) {