Studio: render figure regions (vector + raster) for RAG captioning

page.get_images() only returns raster blobs embedded in the PDF's
resource dictionary, so vector schematics like Figure 1 — drawn purely
with paths/lines — were never extracted, and the VLM only ever saw
incidental embedded photos that happened to live near figures.

Replace the xref-based extraction with bbox rendering: union the
bounding rects of all vector drawings and raster image_info entries on
each page, expand a few points, and render the region with
get_pixmap(clip=bbox, matrix=2x). The captioner now receives the
actual figure — schematic arrows, box labels, legend text, and any
inset photos — and produces a caption that describes the figure as a
whole, not just one embedded sub-image.

Also sharpen the captioner prompt: explicitly tell the VLM the image
is a single figure cropped from a PDF page, and not to describe page
chrome or body paragraphs.

studio/backend/core/rag/captioner.py

@@ -33,9 +33,14 @@ from loggers import get_logger
 logger = get_logger(__name__)
 
 _PROMPT = (
-    "Describe this figure in <=60 words. Focus on factual content "
-    "(axes, labels, captions, visible text, main objects). "
-    "Do not speculate beyond what is visible."
+    "This image is a region cropped from a PDF page that contains a "
+    "single figure (schematic, chart, diagram, table, photo, or "
+    "their combination). Describe the figure's structure and content "
+    "in <=80 words. Focus on factual visible content: axes, labels, "
+    "arrow labels, box labels, legends, visible text in the figure, "
+    "and what entities are connected to what. Do not speculate beyond "
+    "what is visible and do not describe the page header/footer or "
+    "body paragraphs."
 )
 _MAX_NEW_TOKENS = 200
 # Downscale large images so the base64 payload stays manageable; the chat

studio/backend/core/rag/parsers/pdf.py

@@ -58,45 +58,76 @@ def _extract_with_pymupdf(path: Path, want_images: bool) -> ParseResult:
         doc.close()
 
 
+# Pages smaller than this (in PDF points) are ignored as figure regions —
+# bigger than a typical icon/glyph, smaller than a banner.
+_MIN_FIGURE_PT = 60
+# 2× scale renders at 144 dpi (PDF default is 72 dpi). Enough resolution for
+# the captioner to read axis labels, arrow text, and inset photos.
+_RENDER_SCALE = 2.0
+# Expand the union bbox a few points so caption baselines / borders survive.
+_FIGURE_MARGIN_PT = 8.0
+
+
 def _extract_images_pymupdf(doc, pages: list[ParsedPage]) -> list[ParsedImage]:
+    """Render each page's figure region (vector drawings + raster sub-images)
+    as a single PNG. Vector schematics like Figure 1 (no embedded raster)
+    are visible to the captioner only via rendering — ``page.get_images``
+    misses them entirely. We union all non-text geometry on a page into
+    one bbox; for academic papers this typically maps 1:1 to "the figure
+    on this page".
+    """
+    import pymupdf
+
     captions_by_page: dict[int, str] = {
         p.page_number: p.text for p in pages if p.page_number
     }
     out: list[ParsedImage] = []
     for page_index in range(len(doc)):
+        page = doc[page_index]
         page_number = page_index + 1
         try:
-            image_list = doc[page_index].get_images(full = True)
+            rects: list[pymupdf.Rect] = []
+            for drawing in page.get_drawings() or []:
+                rect = drawing.get("rect")
+                if rect is not None:
+                    rects.append(pymupdf.Rect(rect))
+            for info in page.get_image_info(xrefs = True) or []:
+                bbox = info.get("bbox")
+                if bbox is not None:
+                    rects.append(pymupdf.Rect(bbox))
         except Exception:
             continue
-        for img_info in image_list:
-            xref = img_info[0]
-            try:
-                extracted = doc.extract_image(xref)
-            except Exception:
-                continue
-            image_bytes = extracted.get("image")
-            ext = (extracted.get("ext") or "png").lower()
-            mime = {
-                "png": "image/png",
-                "jpg": "image/jpeg",
-                "jpeg": "image/jpeg",
-                "gif": "image/gif",
-                "webp": "image/webp",
-                "bmp": "image/bmp",
-                "tiff": "image/tiff",
-            }.get(ext, f"image/{ext}")
-            if not image_bytes:
-                continue
-            caption = (captions_by_page.get(page_number, "") or "")[:1500]
-            out.append(
-                ParsedImage(
-                    image_bytes = image_bytes,
-                    mime_type = mime,
-                    page_number = page_number,
-                    nearest_caption = caption,
-                )
+        if not rects:
+            continue
+        union = rects[0]
+        for r in rects[1:]:
+            union |= r
+        if union.width < _MIN_FIGURE_PT or union.height < _MIN_FIGURE_PT:
+            continue
+        # Expand and clip to page rect so we don't render past page edges.
+        union = pymupdf.Rect(
+            union.x0 - _FIGURE_MARGIN_PT,
+            union.y0 - _FIGURE_MARGIN_PT,
+            union.x1 + _FIGURE_MARGIN_PT,
+            union.y1 + _FIGURE_MARGIN_PT,
+        ) & page.rect
+        try:
+            matrix = pymupdf.Matrix(_RENDER_SCALE, _RENDER_SCALE)
+            pix = page.get_pixmap(clip = union, matrix = matrix, alpha = False)
+            png_bytes = pix.tobytes("png")
+        except Exception:
+            continue
+        if not png_bytes:
+            continue
+        caption = (captions_by_page.get(page_number, "") or "")[:1500]
+        out.append(
+            ParsedImage(
+                image_bytes = png_bytes,
+                mime_type = "image/png",
+                page_number = page_number,
+                nearest_caption = caption,
             )
+        )
     return out