feat: speed up ultra fast translation

- Batch paragraphs per page through POST /api/translate-batch for
  OpenAI-compatible and ChatGPT Subscription providers, sharing the
  single-paragraph cache and falling back per paragraph on bad replies.
- Adapt per-provider concurrency (AIMD) on HTTP 429, honoring retry-after.
- Let reader tasks carry a page priority so the visible and requested
  pages translate first.
- Prewarm pdf.js once and again on document open to cut first-page latency.

Author: reycn

server/app.mjs

@@ -7,7 +7,7 @@ import { createDocumentCache } from './documents/document-cache.mjs';
 import { createDocumentStore, isDocumentId, MAX_DOCUMENT_BYTES } from './documents/documents.mjs';
 import { createCompanionDocumentStore } from './documents/companion-document-store.mjs';
 import { createLayoutExtraction } from './documents/layout-extraction.mjs';
-import { extractTextWithPositionsAsync } from './documents/pdf-extractor.mjs';
+import { extractTextWithPositionsAsync, prewarmPdfJs } from './documents/pdf-extractor.mjs';
 import { createPerformanceTracker } from './diagnostics/performance.mjs';
 import { createDeveloperDiagnostics } from './diagnostics/developer-diagnostics.mjs';
 import {
@@ -53,6 +53,7 @@ import { registerProviderModelRoutes } from './http/provider-model-routes.mjs';
 import { registerReaderTaskRoutes } from './http/reader-task-routes.mjs';
 import { createFormulaOcrService } from './formula/formula-ocr-service.mjs';
 import { createReaderTasks } from './translation/reader-tasks.mjs';
+import { createAdaptiveLimits } from './translation/adaptive-limits.mjs';
 
 export async function startServer({
   port = 5173,
@@ -121,6 +122,7 @@ export async function startServer({
   const documentCache = createDocumentCache(cacheDir);
   const translationVersions = createTranslationVersions({ directory: cacheDir, documentCache });
   const providerLimiter = createLimiter(4, { label: 'provider' });
+  const adaptiveLimits = createAdaptiveLimits();
   const pageLimiter = createLimiter(2, { label: 'kernel-pages' });
   const proxyJobs = new Map();
   const kernelReports = [];
@@ -173,6 +175,7 @@ export async function startServer({
     isBusy: () =>
       proxyJobs.size > 0 ||
       providerLimiter.snapshot().tasks.length > 0 ||
+      adaptiveLimits.snapshots().some((snapshot) => snapshot.tasks.length > 0) ||
       pageLimiter.snapshot().tasks.length > 0 ||
       (engines.tasks?.().length || 0) > 0 ||
       (readerTasks?.active?.() || 0) > 0,
@@ -188,7 +191,11 @@ export async function startServer({
     proxyJobs,
     proxyUrl: () => `${origin}/kernel-proxy/v1`,
   });
-  const diagnosticQueues = () => [pageLimiter.snapshot(), providerLimiter.snapshot()];
+  const diagnosticQueues = () => [
+    pageLimiter.snapshot(),
+    providerLimiter.snapshot(),
+    ...adaptiveLimits.snapshots(),
+  ];
   const diagnosticTasks = () => {
     const tasks = [...(engines.tasks?.() || [])];
     for (const queue of diagnosticQueues()) tasks.push(...queue.tasks);
@@ -252,6 +259,7 @@ export async function startServer({
     maxDocumentBytes: MAX_DOCUMENT_BYTES,
     translationVersions,
     onDocumentDelete: (id) => readerTasks.cancelForDocument(id),
+    onDocumentOpen: () => void prewarmPdfJs(),
   });
   registerKernelRoutes(app, {
     engines,
@@ -288,6 +296,7 @@ export async function startServer({
     providerRuntime,
     getApiKey,
     translationVersions,
+    adaptiveLimits,
   });
   registerReaderTaskRoutes(app, { readerTasks });
 

server/documents/pdf-extractor.mjs

@@ -10,8 +10,8 @@ export function hasNativeInspector(platform = process.platform, arch = process.a
       : platform === 'linux' && ['x64', 'arm64'].includes(arch);
 }
 
-let native, pdfjs, recoveryWarned;
-export async function extractPdfJsPositions(bytes, pages) {
+let native, pdfjs, recoveryWarned, prewarm;
+function loadPdfJs() {
   pdfjs ??= (async () => {
     if (!globalThis.DOMMatrix) globalThis.DOMMatrix = (await import('@thednp/dommatrix')).default;
     const module = await import('pdfjs-dist/legacy/build/pdf.mjs');
@@ -22,7 +22,45 @@ export async function extractPdfJsPositions(bytes, pages) {
     ).href;
     return { module, root };
   })();
-  const { module, root } = await pdfjs;
+  return pdfjs;
+}
+
+function blankPdf() {
+  const objects = [
+    '<< /Type /Catalog /Pages 2 0 R >>',
+    '<< /Type /Pages /Kids [3 0 R] /Count 1 >>',
+    '<< /Type /Page /Parent 2 0 R /MediaBox [0 0 10 10] >>',
+  ];
+  let body = '%PDF-1.4\n';
+  const offsets = objects.map((object, index) => {
+    const offset = body.length;
+    body += `${index + 1} 0 obj\n${object}\nendobj\n`;
+    return offset;
+  });
+  const xref = body.length;
+  body += `xref\n0 ${objects.length + 1}\n0000000000 65535 f \n`;
+  for (const offset of offsets) body += `${String(offset).padStart(10, '0')} 00000 n \n`;
+  body += `trailer\n<< /Size ${objects.length + 1} /Root 1 0 R >>\nstartxref\n${xref}\n%%EOF\n`;
+  return Buffer.from(body, 'latin1');
+}
+
+// Scanned-page recovery cold-loads pdf.js and its fake worker on first use,
+// which costs most of a second. Warm it in the background once a document is
+// open so the first scanned page does not pay that cost on the reading path.
+export function prewarmPdfJs({ extract = extractPdfJsPositions } = {}) {
+  prewarm ??= new Promise((resolveWarm) => {
+    const timer = setTimeout(() => {
+      extract(blankPdf(), [1])
+        .catch(() => {})
+        .then(() => resolveWarm());
+    }, 0);
+    timer.unref?.();
+  });
+  return prewarm;
+}
+
+export async function extractPdfJsPositions(bytes, pages) {
+  const { module, root } = await loadPdfJs();
   const task = module.getDocument({
     data: new Uint8Array(bytes),
     isEvalSupported: false,

server/http/document-routes.mjs

@@ -64,6 +64,7 @@ export function registerDocumentRoutes(
     maxDocumentBytes,
     translationVersions,
     onDocumentDelete,
+    onDocumentOpen,
   },
 ) {
   app.post(
@@ -79,6 +80,7 @@ export function registerDocumentRoutes(
           .register(entry.documentHash, req.headers['x-document-name'])
           .catch(() => {});
         res.status(201).json({ id });
+        onDocumentOpen?.(id);
       } catch (e) {
         if (e?.status) return res.status(e.status).json({ error: e.message });
         res.status(400).json({ error: e.message });