feat(docling): detect response regions with OpenCV

tests/test_docling_regions.py

@@ -0,0 +1,39 @@
+from __future__ import annotations
+
+from PIL import Image, ImageDraw
+
+from api.services.docling.regions import detect_response_regions_from_image
+
+
+def test_detects_grouped_answer_lines() -> None:
+    image = Image.new("RGB", (900, 1200), "white")
+    draw = ImageDraw.Draw(image)
+    for y in (420, 470, 520):
+        draw.line((160, y, 760, y), fill="black", width=3)
+
+    candidates = detect_response_regions_from_image(image, page_index=2)
+
+    line_regions = [c.to_mapper_dict() for c in candidates if c.region_type == "answer_lines"]
+    assert line_regions
+    best = line_regions[0]
+    assert best["kind"] == "response"
+    assert best["source"] == "ai"
+    assert best["confirmed"] is False
+    assert best["page_index"] == 2
+    assert best["line_count"] == 3
+    assert best["bbox"]["coord_origin"] == "TOPLEFT"
+    assert best["bbox"]["w"] > 550
+    assert best["bbox"]["h"] > 80
+
+
+def test_detects_answer_box() -> None:
+    image = Image.new("RGB", (900, 1200), "white")
+    draw = ImageDraw.Draw(image)
+    draw.rectangle((140, 300, 780, 520), outline="black", width=3)
+
+    candidates = detect_response_regions_from_image(image, page_index=0)
+
+    boxes = [c.to_mapper_dict() for c in candidates if c.region_type == "answer_box"]
+    assert boxes
+    assert boxes[0]["bbox"]["w"] > 600
+    assert boxes[0]["bbox"]["h"] > 200