Updated Upstage OCR Normalization Logic

tasks.py

@@ -109,37 +109,52 @@ def parse_ocr_text(self, presigned_url: str, request_id: str, file_name: str):
 
 
 # Upstage 응답 정규화: 가능한 많은 'text'를 모으고, 후보 bbox를 수집
-def _normalize_upstage_response(resp_json):
+def _normalize_upstage_response(resp_json, return_word_level=False, normalize=False):
     """
     Upstage 문서 디지타이제이션 응답에서 text와 bbox 후보를 추출.
     구조가 달라도 dict/list를 재귀 탐색하여 'text' 유사 키와 bbox 유사 키를 모읍니다.
     """
-    texts = []
-    boxes = []
+    # 1) 전체 텍스트 추출
+    if isinstance(resp_json, dict) and resp_json.get("text"):
+        full_text = resp_json["text"]
+    else:
+        full_text = ""
+        for p in resp_json.get("pages") or []:
+            t = p.get("text")
+            if t:
+                full_text += t + "\n"
+        full_text = full_text.rstrip("\n")
 
-    def walk(obj):
-        if isinstance(obj, dict):
-            for k, v in obj.items():
-                kl = k.lower()
-                # text 후보 키
-                if kl in ("text", "content", "ocr_text", "full_text", "value"):
-                    if isinstance(v, str) and v.strip():
-                        texts.append(v.strip())
-                # bbox/box 후보 키
-                if kl in ("bbox", "box", "bounding_box", "boundingbox", "polygon"):
-                    boxes.append(v)
-                # 재귀
-                walk(v)
-        elif isinstance(obj, list):
-            for item in obj:
-                walk(item)
+    # 2) 좌표/워드 추출
+    coords = []
+    word_items = []
+    pages = resp_json.get("pages") or []
+    for p_idx, page in enumerate(pages, start=1):
+        w = page.get("words") or []
+        pw, ph = page.get("width"), page.get("height")  # 정규화 옵션용
+        for wobj in w:
+            bb = (wobj.get("boundingBox") or {}).get("vertices") or []
+            if len(bb) == 4:
+                poly = [[float(pt.get("x", 0)), float(pt.get("y", 0))] for pt in bb]
 
-    walk(resp_json)
+                if normalize and pw and ph:
+                    poly = [[x / float(pw), y / float(ph)] for x, y in poly]
 
-    merged_text = (
-        "\n".join(texts) if texts else json.dumps(resp_json, ensure_ascii=False)
-    )
-    return merged_text, boxes
+                if return_word_level:
+                    word_items.append(
+                        {
+                            "page": p_idx,
+                            "text": wobj.get("text", ""),
+                            "confidence": float(wobj.get("confidence") or 0.0),
+                            "box": poly,  # 4x2
+                        }
+                    )
+                else:
+                    coords.append(poly)
+
+    if return_word_level:
+        return full_text, word_items
+    return full_text, coords
 
 
 # (Upstage) 외부 OCR API 호출 + 후처리