@tiens.nguyen/gonext-local-worker 1.0.101 → 1.0.103

package/gonext-local-worker.mjs

@@ -1052,6 +1052,42 @@ async function correctOcrTextViaServer(extractedText, baseUrl) {
   return normalizeCorrection(corrected, extractedText);
 }
 
+/**
+ * Call an Ollama host's native chat endpoint (POST {base}/api/chat) for text
+ * correction/translation, e.g. base "https://ollama1.gomarsic.cc" model "qwen3:14b".
+ * `think:false` suppresses reasoning models' <think> preamble so we get clean text.
+ */
+async function correctOcrTextViaOllama(extractedText, base, model) {
+  const systemPrompt =
+    "You are a text correction assistant. Fix grammar and language errors in the " +
+    "provided text while preserving the original meaning and language. " +
+    "Return only the corrected text without any explanation.";
+  const res = await fetch(`${base.replace(/\/+$/, "")}/api/chat`, {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({
+      model,
+      stream: false,
+      think: false,
+      options: { temperature: 0 },
+      messages: [
+        { role: "system", content: systemPrompt },
+        { role: "user", content: `Text:\n${extractedText}` },
+      ],
+    }),
+    signal: AbortSignal.timeout(OCR_CORRECT_TIMEOUT_MS),
+  });
+  if (!res.ok) {
+    throw new Error(`Ollama ${base} returned ${res.status}`);
+  }
+  const json = await res.json();
+  let corrected = json?.message?.content?.trim() || "";
+  // Safety net for reasoning models (qwen3) if `think:false` isn't honored:
+  // drop any <think>…</think> preamble before normalizing.
+  corrected = corrected.replace(/<think>[\s\S]*?<\/think>/gi, "").trim();
+  return normalizeCorrection(corrected, extractedText);
+}
+
 async function correctOcrText(extractedText, modelOverride = "") {
   if (OCR_SKIP_CORRECTION) {
     console.log("[gonext-worker] OCR correction skipped (GONEXT_OCR_SKIP_CORRECTION=1)");
@@ -1060,6 +1096,38 @@ async function correctOcrText(extractedText, modelOverride = "") {
 
   const override = modelOverride.trim();
 
+  // Ollama-backed correction/translation, e.g. "ollama:qwen3:14b@@https://ollama1.gomarsic.cc".
+  // The model may itself contain a colon (qwen3:14b), so only strip the leading
+  // "ollama:" scheme and split the base off on the "@@" delimiter.
+  if (override.toLowerCase().startsWith("ollama:")) {
+    const rest = override.slice("ollama:".length);
+    const at = rest.indexOf("@@");
+    const model = (at >= 0 ? rest.slice(0, at) : rest).trim();
+    const base = (at >= 0 ? rest.slice(at + 2) : "").trim().replace(/\/+$/, "");
+    if (!model || !base) {
+      console.warn(
+        `[gonext-worker] OCR correction: malformed ollama override "${override}" (using raw OCR text)`
+      );
+      return extractedText;
+    }
+    console.log(`[gonext-worker] OCR correction via Ollama ${base} model=${model}`);
+    console.log(`[gonext-worker] OCR correction input: ${extractedText.slice(0, 300)}`);
+    try {
+      const corrected = await correctOcrTextViaOllama(extractedText, base, model);
+      console.log(`[gonext-worker] OCR correction output: ${corrected.slice(0, 300)}`);
+      console.log(
+        `[gonext-worker] OCR correction done: ${extractedText.length} → ${corrected.length} chars`
+      );
+      return corrected;
+    } catch (e) {
+      const msg = e instanceof Error ? e.message : String(e);
+      console.warn(
+        `[gonext-worker] OCR correction Ollama failed (using raw OCR text): ${msg.slice(0, 200)}`
+      );
+      return extractedText;
+    }
+  }
+
   // If the override looks like a URL, call it as a running mlx_lm.server endpoint.
   // e.g. "http://127.0.0.1:8083/v1" or "http://127.0.0.1:8083"
   if (override.startsWith("http://") || override.startsWith("https://")) {
@@ -1685,7 +1753,12 @@ function logModelResponseToWorker(jobId, modelId, text) {
 async function checkOllamaTags(base) {
   const endpoint = `${base}/api/tags`;
   try {
-    const res = await fetch(endpoint, { method: "GET" });
+    // Bounded so an unreachable/slow host fails fast (a hung fetch would exceed
+    // the client's local-health wait and leave the status dot stuck "pending").
+    const res = await fetch(endpoint, {
+      method: "GET",
+      signal: AbortSignal.timeout(4000),
+    });
     if (!res.ok) return { online: false, endpoint, models: [] };
     const j = await res.json();
     const source = sourceLabelFromBase(base);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.101",
+  "version": "1.0.103",
   "description": "Polls GoNext cloud API for async local LLM jobs and runs them against Ollama/OpenAI-compatible servers on this Mac",
   "type": "module",
   "license": "MIT",