🧹 [refactor-perf] Optimize engine thread pool, event loop, and resultobj initialization

src/api/main.py

@@ -241,8 +241,8 @@ async def openai_vision_endpoint(
     try:
         # Remove data URI prefix if present
         header, encoded = img_data.split(",", 1) if "," in img_data else (None, img_data)
-        contents = base64.b64decode(encoded)
-        img = Image.open(io.BytesIO(contents))
+        contents = await asyncio.to_thread(base64.b64decode, encoded)
+        img = await asyncio.to_thread(Image.open, io.BytesIO(contents))
 
         # Dimension check
         if img.width * img.height > MAX_PIXELS:
@@ -321,7 +321,7 @@ async def _get_image_from_request(request: Request, file: Optional[UploadFile]):
                 raise HTTPException(status_code=499, detail="Client closed connection")
             if len(contents) > MAX_IMAGE_SIZE:
                 raise HTTPException(status_code=413, detail="File too large")
-            img = Image.open(io.BytesIO(contents))
+            img = await asyncio.to_thread(Image.open, io.BytesIO(contents))
             filename = file.filename or "uploaded_image.jpg"
         else:
             # Handle JSON body (Base64)
@@ -346,8 +346,8 @@ async def _get_image_from_request(request: Request, file: Optional[UploadFile]):
             ocr_req = OCRRequest(**body)
             # Remove data URI prefix if present
             header, encoded = ocr_req.image.split(",", 1) if "," in ocr_req.image else (None, ocr_req.image)
-            contents = base64.b64decode(encoded)
-            img = Image.open(io.BytesIO(contents))
+            contents = await asyncio.to_thread(base64.b64decode, encoded)
+            img = await asyncio.to_thread(Image.open, io.BytesIO(contents))
             filename = "base64_image.jpg"
     except HTTPException:
         raise

src/core/engine.py

@@ -86,8 +86,17 @@ def __init__(
         self.recognizer50 = None
 
         # ThreadPoolExecutor for parallelizing character recognition across lines
+        # Allow override via environment variable OCR_MAX_WORKERS, or use a bounded default.
+        # Too many concurrent threads calling ORT sessions on CPU causes high contention and thrashing.
+        max_workers_env = os.getenv("OCR_MAX_WORKERS")
+        if max_workers_env:
+            max_workers = int(max_workers_env)
+        else:
+            cpu_count = os.cpu_count() or 4
+            max_workers = min(4, cpu_count)
+
         self.executor = ThreadPoolExecutor(
-            max_workers=os.cpu_count() or 4,
+            max_workers=max_workers,
             thread_name_prefix="ocr_worker"
         )
 
@@ -221,14 +230,14 @@ def ocr(self, pil_image: Image.Image, img_name: str = "image.jpg") -> Dict[str,
         classeslist = list(self.detector.classes.values())
 
         # Prepare data for NDL-style XML conversion
-        resultobj = [dict(), dict()]
-        resultobj[0][0] = list()
-        for i in range(17):
-            resultobj[1][i] = []
+        resultobj = [
+            {0: []},
+            {i: [] for i in range(17)}
+        ]
         for det in detections:
             xmin, ymin, xmax, ymax = det["box"]
             conf = det["confidence"]
-            char_count = det.get("pred_char_count", 100.0) # v1.2.1 uses char_count in resultobj
+            char_count = det.get("pred_char_count", 100.0)  # v1.2.1 uses char_count in resultobj
             if det["class_index"] == 0:
                 resultobj[0][0].append([xmin, ymin, xmax, ymax])
             # v1.2.1 adds char_count here