@tiens.nguyen/gonext-local-worker 1.0.108 → 1.0.109

package/gonext-local-worker.mjs

@@ -139,6 +139,10 @@ const OCR_TIMEOUT_MS = 180_000;
 const OCR_DEBUG = String(process.env.GONEXT_OCR_DEBUG ?? "").trim() === "1";
 const OCR_SKIP_CORRECTION = String(process.env.GONEXT_OCR_SKIP_CORRECTION ?? "").trim() === "1";
 const OCR_CORRECT_TIMEOUT_MS = 120_000;
+// OCR translation/correction is hardcoded to a managed Ollama service + model.
+// (The user-configurable "OCR translation model" chooser was removed.)
+const OCR_TRANSLATE_OLLAMA_URL = "https://ollama1.gomarsic.cc";
+const OCR_TRANSLATE_OLLAMA_MODEL = "gemma4:26b";
 
 async function workerFetch(path, init = {}) {
   const url = `${apiBase}${path.startsWith("/") ? path : `/${path}`}`;
@@ -987,16 +991,6 @@ function normalizeOcrOutput(output) {
   return lines.join("\n").trim();
 }
 
-function resolveOcrCorrectionModelPath(override) {
-  if (override) {
-    const expanded = override.replace(/^~(?=\/)/, homedir());
-    return expanded.startsWith("/") ? expanded : join(homedir(), "mlx-models", expanded);
-  }
-  const fromEnv = String(process.env.GONEXT_OCR_CORRECT_MODEL_PATH ?? "").trim();
-  if (fromEnv) return fromEnv.replace(/^~(?=\/)/, homedir());
-  return join(homedir(), "mlx-models", "translategemma-4b-it-4bit");
-}
-
 function normalizeCorrection(raw, originalText) {
   let text = String(raw ?? "").replace(/\r\n/g, "\n");
   // Strip common mlx_lm.generate noise lines
@@ -1020,38 +1014,6 @@ function normalizeCorrection(raw, originalText) {
   return result;
 }
 
-/**
- * Call a correction model that's already running as an OpenAI-compatible server
- * (e.g. mlx_lm.server --model translategemma-4b-it-4bit --port 8083).
- * baseUrl should be like http://127.0.0.1:8083/v1
- */
-async function correctOcrTextViaServer(extractedText, baseUrl) {
-  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(`${baseUrl.replace(/\/+$/, "")}/chat/completions`, {
-    method: "POST",
-    headers: { "Content-Type": "application/json" },
-    body: JSON.stringify({
-      model: "default",
-      messages: [
-        { role: "system", content: systemPrompt },
-        { role: "user", content: `Text:\n${extractedText}` },
-      ],
-      max_tokens: 2048,
-      temperature: 0,
-    }),
-    signal: AbortSignal.timeout(OCR_CORRECT_TIMEOUT_MS),
-  });
-  if (!res.ok) {
-    throw new Error(`Correction server ${baseUrl} returned ${res.status}`);
-  }
-  const json = await res.json();
-  const corrected = json?.choices?.[0]?.message?.content?.trim() || "";
-  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".
@@ -1088,108 +1050,35 @@ async function correctOcrTextViaOllama(extractedText, base, model) {
   return normalizeCorrection(corrected, extractedText);
 }
 
-async function correctOcrText(extractedText, modelOverride = "") {
+async function correctOcrText(extractedText) {
   if (OCR_SKIP_CORRECTION) {
     console.log("[gonext-worker] OCR correction skipped (GONEXT_OCR_SKIP_CORRECTION=1)");
     return extractedText;
   }
 
-  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): ${msg.slice(0, 200)}`
-      );
-      return extractedText;
-    }
-  }
-
-  // If the override looks like a URL, call it as a running mlx_lm.server endpoint.
-  // e.g. "http://127.0.0.1:8083/v1" or "http://127.0.0.1:8083"
-  if (override.startsWith("http://") || override.startsWith("https://")) {
-    const baseUrl = /\/v1\/?$/i.test(override) ? override : `${override}/v1`;
-    console.log(`[gonext-worker] OCR correction via server ${baseUrl}`);
-    console.log(`[gonext-worker] OCR correction input: ${extractedText.slice(0, 300)}`);
-    try {
-      const corrected = await correctOcrTextViaServer(extractedText, baseUrl);
-      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 server failed (using raw OCR text): ${msg.slice(0, 200)}`);
-      return extractedText;
-    }
-  }
-
-  // Otherwise invoke mlx_lm.generate CLI (model loads per job — no server needed).
-  const modelPath = resolveOcrCorrectionModelPath(override);
-  const prompt =
-    "The following text was extracted by an OCR model and may contain grammar errors, " +
-    "wrong language, or garbled characters. " +
-    "Correct any errors while preserving the original meaning and language. " +
-    "Return only the corrected text without any explanation.\n\n" +
-    `Text:\n${extractedText}`;
-  console.log(`[gonext-worker] OCR correction via CLI model=${modelPath}`);
+  // OCR translation/correction is hardcoded to a managed Ollama service + model
+  // (the user-configurable chooser was removed). Any error falls back to raw text.
+  console.log(
+    `[gonext-worker] OCR correction via Ollama ${OCR_TRANSLATE_OLLAMA_URL} model=${OCR_TRANSLATE_OLLAMA_MODEL}`
+  );
   console.log(`[gonext-worker] OCR correction input: ${extractedText.slice(0, 300)}`);
-  // Write prompt to a temp Python script that calls mlx_lm programmatically,
-  // avoiding CLI arg length/escaping limits and the missing --prompt-file flag.
-  const scriptFile = join(tmpdir(), `gonext-ocr-correct-${Date.now()}.py`);
-  const promptFile = join(tmpdir(), `gonext-ocr-correct-${Date.now()}.txt`);
-  const pyScript = `import sys, mlx_lm
-model, tokenizer = mlx_lm.load(sys.argv[1])
-with open(sys.argv[2], encoding="utf-8") as f:
-    prompt = f.read()
-result = mlx_lm.generate(model, tokenizer, prompt=prompt, max_tokens=2048, verbose=False)
-print(result)
-`;
   try {
-    await writeFile(scriptFile, pyScript, "utf8");
-    await writeFile(promptFile, prompt, "utf8");
-    const { stdout, stderr } = await execFile(
-      "python3",
-      [scriptFile, modelPath, promptFile],
-      { timeout: OCR_CORRECT_TIMEOUT_MS, maxBuffer: 10 * 1024 * 1024 }
+    const corrected = await correctOcrTextViaOllama(
+      extractedText,
+      OCR_TRANSLATE_OLLAMA_URL,
+      OCR_TRANSLATE_OLLAMA_MODEL
     );
-    if (stderr?.trim()) {
-      console.log(`[gonext-worker] OCR correction CLI stderr: ${stderr.trim().slice(0, 300)}`);
-    }
-    const corrected = normalizeCorrection(stdout, extractedText);
     console.log(`[gonext-worker] OCR correction output: ${corrected.slice(0, 300)}`);
-    console.log(`[gonext-worker] OCR correction done: ${extractedText.length} → ${corrected.length} chars`);
+    console.log(
+      `[gonext-worker] OCR correction done: ${extractedText.length} → ${corrected.length} chars`
+    );
     return corrected;
   } catch (e) {
-    const stderr = e && typeof e === "object" && "stderr" in e ? String(e.stderr ?? "").trim().slice(0, 300) : "";
     const msg = e instanceof Error ? e.message : String(e);
-    console.warn(`[gonext-worker] OCR correction CLI failed (using raw OCR text): ${msg.slice(0, 200)}${stderr ? ` | stderr: ${stderr}` : ""}`);
+    console.warn(
+      `[gonext-worker] OCR correction Ollama failed (using raw OCR text): ${msg.slice(0, 200)}`
+    );
     return extractedText;
-  } finally {
-    await rm(scriptFile, { force: true }).catch(() => {});
-    await rm(promptFile, { force: true }).catch(() => {});
   }
 }
 
@@ -1335,13 +1224,10 @@ async function runOcrJob(job) {
     } finally {
       await rm(tempDir, { recursive: true, force: true }).catch(() => {});
     }
-    // Post-process: correct grammar/language with translategemma-4b-it-4bit
-    // (or the model set by the user in Settings > OCR correction model).
-    // Falls back to raw OCR text on any error so the job never fails here.
-    const correctionModelOverride = typeof payload.correctionModel === "string"
-      ? payload.correctionModel.trim()
-      : "";
-    extractedText = await correctOcrText(extractedText, correctionModelOverride);
+    // Post-process: correct grammar/language via the hardcoded managed Ollama
+    // service (gemma4:26b). Falls back to raw OCR text on any error so the job
+    // never fails here. (`payload.correctionModel` is no longer used.)
+    extractedText = await correctOcrText(extractedText);
     const totalTimeSeconds = (Date.now() - start) / 1000;
     const doneRes = await workerFetch(`/api/worker/jobs/${jobId}`, {
       method: "PATCH",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.108",
+  "version": "1.0.109",
   "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",