Update app.py

app.py

@@ -146,9 +146,15 @@ def run_ocr(pil_image: Image.Image, mode: str = "free") -> str:
 
     try:
         if hasattr(model, "infer"):
+            # ── Strategy 1: capture stdout ──────────────────────────────────
+            # model.infer() prints the OCR result to stdout instead of returning it.
+            # We capture stdout + also try save_results=True as backup.
+            import io, sys
+            from contextlib import redirect_stdout
+
             with tempfile.TemporaryDirectory() as out_dir:
-                # force_cpu() patches .cuda() → no-op so model.infer() works on CPU
-                with force_cpu():
+                stdout_buf = io.StringIO()
+                with force_cpu(), redirect_stdout(stdout_buf):
                     result = model.infer(
                         tokenizer,
                         prompt=f"<image>\n{prompt_text}",
@@ -157,11 +163,53 @@ def run_ocr(pil_image: Image.Image, mode: str = "free") -> str:
                         base_size=1024,
                         image_size=768,
                         crop_mode=True,
-                        save_results=False,
+                        save_results=True,   # also write to file as backup
                     )
-            if isinstance(result, dict):
-                return result.get("text", str(result))
-            return str(result) if result else ""
+
+                # Echo captured stdout to real stdout for server logs
+                captured = stdout_buf.getvalue()
+                sys.stdout.write(captured)
+                sys.stdout.flush()
+
+                # ── Extract text: priority order ─────────────────────────────
+                text = ""
+
+                # 1) Return value (if model returns text directly)
+                if result:
+                    if isinstance(result, dict):
+                        text = result.get("text", result.get("output", ""))
+                    elif isinstance(result, str):
+                        text = result
+
+                # 2) Captured stdout (most reliable for this model)
+                if not text and captured:
+                    # The model prints: "===================== <actual text>"
+                    # Strip the separator and any leading/trailing whitespace
+                    cleaned = captured.strip()
+                    for sep in ["=====================", "=====", "-----"]:
+                        if sep in cleaned:
+                            cleaned = cleaned.split(sep, 1)[-1].strip()
+                            break
+                    text = cleaned
+
+                # 3) Output files written by save_results=True
+                if not text:
+                    import glob
+                    for ext in ["*.txt", "*.md", "*.json"]:
+                        files = glob.glob(os.path.join(out_dir, "**", ext), recursive=True)
+                        for fpath in files:
+                            try:
+                                with open(fpath, "r", encoding="utf-8") as f:
+                                    file_text = f.read().strip()
+                                if file_text:
+                                    text = file_text
+                                    break
+                            except Exception:
+                                pass
+                        if text:
+                            break
+
+            return text
 
         # ── Fallback: standard generate() if model.infer() is not available ──
         messages = [{"role": "user", "content": [