@tiens.nguyen/gonext-local-worker 1.0.220 → 1.0.222

package/gonext-local-worker.mjs

@@ -1249,8 +1249,15 @@ async function correctOcrTextViaOpenAI(extractedText, base, model) {
       model,
       stream: false,
       temperature: 0,
+      // Qwen3 (and other reasoning models) emit a <think> trace by default. Over the
+      // OpenAI /v1 API there's no `think:false` flag, so use Qwen3's `/no_think` SOFT
+      // SWITCH (honored by its chat template, same as the agent loop) to suppress it —
+      // otherwise the model spends its whole token budget reasoning and returns no
+      // corrected text (seen live: correction was a no-op, chars unchanged). Generous
+      // max_tokens so a long OCR page isn't truncated.
+      max_tokens: 4096,
       messages: [
-        { role: "system", content: OCR_CORRECT_SYSTEM_PROMPT },
+        { role: "system", content: `${OCR_CORRECT_SYSTEM_PROMPT}\n/no_think` },
         { role: "user", content: `Text:\n${extractedText}` },
       ],
     }),
@@ -1261,7 +1268,13 @@ async function correctOcrTextViaOpenAI(extractedText, base, model) {
   }
   const json = await res.json();
   let corrected = json?.choices?.[0]?.message?.content?.trim() || "";
-  corrected = corrected.replace(/<think>[\s\S]*?<\/think>/gi, "").trim();
+  // Strip a <think> trace if `/no_think` wasn't honored: closed pairs, a template-opened
+  // leading …</think>, and a trailing UNCLOSED <think> (truncated by max_tokens).
+  corrected = corrected
+    .replace(/<think>[\s\S]*?<\/think>/gi, "")
+    .replace(/^[\s\S]*?<\/think>/i, "")
+    .replace(/<think>[\s\S]*$/i, "")
+    .trim();
   return normalizeCorrection(corrected, extractedText);
 }
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.220",
+  "version": "1.0.222",
   "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",