@tiens.nguyen/gonext-local-worker 1.0.40 → 1.0.41

package/gonext-local-worker.mjs

@@ -531,7 +531,22 @@ function resolveOcrModelPath() {
 }
 
 function normalizeOcrOutput(output) {
-  const lines = String(output ?? "")
+  let text = String(output ?? "").replace(/\r\n/g, "\n");
+  const afterAssistant = text.includes("<|assistant|>")
+    ? text.split("<|assistant|>").pop()
+    : "";
+  if (afterAssistant && afterAssistant.trim()) {
+    text = afterAssistant;
+  }
+  const escapedPrompt = OCR_PROMPT.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+  text = text
+    .replace(/<\|[^|>]+\|>/g, " ")
+    .replace(/<think>[\s\S]*?<\/think>/gi, " ")
+    .replace(/<think>|<\/think>/gi, " ")
+    .replace(new RegExp(escapedPrompt, "gi"), " ")
+    .replace(/(?:^|[\s/\\])nothink\b/gi, " ")
+    .replace(/\\n/g, "\n");
+  const lines = text
     .split(/\r?\n/)
     .map((line) => line.trim())
     .filter((line) => line.length > 0)
@@ -541,6 +556,7 @@ function normalizeOcrOutput(output) {
         !line.startsWith("Files:") &&
         !line.startsWith("Prompt:") &&
         !line.startsWith("Generation:") &&
+        !line.startsWith("Calling `python -m mlx_vlm.generate") &&
         !line.startsWith("Peak memory:") &&
         !line.startsWith("=======") &&
         line !== "<think>" &&
@@ -551,6 +567,52 @@ function normalizeOcrOutput(output) {
   return lines.join("\n").trim();
 }
 
+async function runMlxVlmGenerate(modelPath, imagePath) {
+  const sharedArgs = [
+    "--model",
+    modelPath,
+    "--prompt",
+    OCR_PROMPT,
+    "--image",
+    imagePath,
+    "--temperature",
+    "0.0",
+    "--max-tokens",
+    String(OCR_MAX_TOKENS),
+  ];
+  try {
+    return await execFile(
+      "python3",
+      ["-m", "mlx_vlm.generate", ...sharedArgs],
+      {
+        timeout: OCR_TIMEOUT_MS,
+        maxBuffer: 10 * 1024 * 1024,
+      }
+    );
+  } catch (primaryError) {
+    const stderr =
+      primaryError && typeof primaryError === "object" && "stderr" in primaryError
+        ? String(primaryError.stderr ?? "").toLowerCase()
+        : "";
+    const message =
+      primaryError instanceof Error ? primaryError.message.toLowerCase() : "";
+    const missingLegacyModule =
+      stderr.includes("no module named mlx_vlm.generate") ||
+      message.includes("no module named mlx_vlm.generate");
+    if (!missingLegacyModule) {
+      throw primaryError;
+    }
+    return execFile(
+      "python3",
+      ["-m", "mlx_vlm", "generate", ...sharedArgs],
+      {
+        timeout: OCR_TIMEOUT_MS,
+        maxBuffer: 10 * 1024 * 1024,
+      }
+    );
+  }
+}
+
 async function runOcrJob(job) {
   const { jobId, payload } = job;
   const start = Date.now();
@@ -587,28 +649,7 @@ async function runOcrJob(job) {
       );
       await writeFile(imagePath, bytes);
       const modelPath = resolveOcrModelPath();
-      const { stdout } = await execFile(
-        "python3",
-        [
-          "-m",
-          "mlx_vlm",
-          "generate",
-          "--model",
-          modelPath,
-          "--prompt",
-          OCR_PROMPT,
-          "--image",
-          imagePath,
-          "--temperature",
-          "0.0",
-          "--max-tokens",
-          String(OCR_MAX_TOKENS),
-        ],
-        {
-          timeout: OCR_TIMEOUT_MS,
-          maxBuffer: 10 * 1024 * 1024,
-        }
-      );
+      const { stdout } = await runMlxVlmGenerate(modelPath, imagePath);
       extractedText = normalizeOcrOutput(stdout);
     } finally {
       await rm(tempDir, { recursive: true, force: true }).catch(() => {});

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.40",
+  "version": "1.0.41",
   "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",