pi-ocr 1.3.0 → 1.3.2

package/extensions/mineru-pro.ts

@@ -144,7 +144,7 @@ async function pollSingleTask(
     const state: string = data.state || "unknown";
 
     if (state === "done") {
-      return await downloadAndExtractMd(data.full_zip_url);
+      return cleanMarkdown(await downloadAndExtractMd(data.full_zip_url));
     }
     if (state === "failed") {
       throw new Error(`MinerU Pro failed: ${data.err_msg || "unknown"}`);
@@ -202,7 +202,7 @@ async function pollBatch(
       for (const r of results) {
         if (r.state === "done" && r.full_zip_url) {
           onProgress(`${progressPrefix} downloading ${r.file_name}…`);
-          const md = await downloadAndExtractMd(r.full_zip_url);
+          const md = cleanMarkdown(await downloadAndExtractMd(r.full_zip_url));
           markdowns.push(md);
         }
       }
@@ -225,6 +225,13 @@ async function pollBatch(
   throw new Error(`MinerU Pro batch ${batchId} timed out`);
 }
 
+// ── Output cleanup ───────────────────────────────────────────────────────────
+
+function cleanMarkdown(md: string): string {
+  // Remove MinerU's embedded image references
+  return md.replace(/!\[.*?\]\(images\/.*?\)\n*/g, "");
+}
+
 // ── Public API ───────────────────────────────────────────────────────────────
 
 export async function mineruProOcr(
@@ -245,5 +252,5 @@ export async function mineruProOcr(
   const markdown = await processLocalFile(token, filePath, fileName, "[1/1]", onProgress);
   onProgress("[1/1] done");
 
-  return { text: markdown, details: { backend: "mineru-pro", fileName } };
+  return { text: cleanMarkdown(markdown), details: { backend: "mineru-pro", fileName } };
 }

package/extensions/pix2text.ts

@@ -46,13 +46,26 @@ def progress(payload):
 
 progress({"status": "loading", "message": "Initializing Pix2Text models..."})
 
+# Auto-detect optimal GPU device
+import torch
+if torch.cuda.is_available():
+    _device = "cuda"
+elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
+    _device = "mps"
+else:
+    _device = "cpu"
+
+# Prevent MPS memory fragmentation on macOS
+if _device == "mps":
+    os.environ.setdefault("PYTORCH_MPS_HIGH_WATERMARK_RATIO", "0.0")
+
 # Suppress model-loading noise on stdout during initialization
 _real_stdout = sys.stdout
 sys.stdout = sys.stderr
 
 try:
     from pix2text import Pix2Text
-    p2t = Pix2Text.from_config(enable_formula=True, enable_table=False)
+    p2t = Pix2Text.from_config(enable_formula=True, enable_table=False, device=_device)
 finally:
     sys.stdout = _real_stdout
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-ocr",
-  "version": "1.3.0",
+  "version": "1.3.2",
   "description": "Pi extension: Zero-setup multi-backend OCR — MinerU (free cloud), Ollama (local GPU, LaTeX formulas), Pix2Text (local Python). Extract text, formulas, and tables from images and PDFs. Default: zero config, works out of the box.",
   "keywords": [
     "pi-package",