CAD建筑图（看图、会审、报告）

概述

CAD Vision Reader

Purpose

The cad-viewer skill can mine text from DWG/DXF, but it cannot see the drawing — graphic geometry, title blocks, dimension lines, and symbols drawn as imagery are invisible to text extraction (especially 天正/Tianzheng tables and scanned signed PDFs).

This skill closes that gap: it renders the final-version drawing PDF into a high-DPI PNG and uses the multimodal Read tool to actually look at it, then outputs a structured six-segment drawing-interpretation report (see references/six_segment_report.md).

When to use

• A drawing is delivered as a signed/scanned PDF (no text layer) and you

need to read the title block, dimensions, or annotations.

• You need to visually verify something that text extraction can't catch

(floor-plan room count, 标高统计表, 工程概况 table, symbol decoding).

• The user asks for a 六段式图纸解读报告 / 图纸会审 / 图纸解读.
• A DWG's text is a 天正 table and cad-viewer returns only fragments.

Dependencies (required)

pip install pymupdf pypdf Pillow

• pymupdf — PDF → PNG rendering (the core of this skill).
• pypdf — optional text probing / page counts.
• Pillow — region cropping from rendered PNGs.

No system packages (poppler/gs/mutool) and no OCR engine are needed: the multimodal Read tool reads the rendered PNG directly.

Optional: DWG text mining (companion, not visual)

For extracting text from .dwg/.dxf (not visual), use the cad-viewer skill's cad_tools.py (needs ezdxf + ODA File Converter). See "DWG handling" below.

Workflow

1. Locate the drawing PDF. Prefer the final / signed version (签章PDF),

which is the authoritative source. Find the per-building PDF (e.g. 建施-12-一层平面图_A2+0.25.pdf) under the project's 签章/审图通过 folder.

1. Render to PNG with the bundled script:

   python {SKILL_DIR}/scripts/render_pdf.py <drawing.pdf> <out.png> --dpi 200
   # region crop (title block) by fractional rect x0,y0,x1,y1 (0..1):
   python {SKILL_DIR}/scripts/render_pdf.py <drawing.pdf> <out.png> --dpi 200 --frac 0.55,0.62,1.0,1.0
   # all pages:
   python {SKILL_DIR}/scripts/render_pdf.py <drawing.pdf> --all <outdir> --dpi 150

Use 200–300 dpi for the full page, 400 dpi if you will crop a small region (title block / 工程概况). Keep each PNG under ~6–8 MB for the Read tool.

1. Look at it with the Read tool on the PNG path(s).
2. Extract title-block fields (图名/图号/比例/版本/日期/设计), dimensions,

annotations, and symbols.

1. Output the six-segment report per references/six_segment_report.md.

DWG handling

• Visual reading of a .dwg is best done by first exporting it to PDF in

your CAD software, then running this skill on the PDF.

• Free ODA File Converter only does DWG/DXF/DGN conversion (no raster PDF),

and ezdxf+matplotlib rendering drops 天正 proxy objects (blank image). So do not rely on direct DWG→image rendering.

• For text inside a DWG, use cad-viewer's cad_tools.py texts / search.

Notes

• The {SKILL_DIR} placeholder = the directory containing this SKILL.md.
• 200 dpi on an A2 sheet ≈ 5900×3300 px (~2.4 MB) — comfortably readable.
• If a region is too small to read at full-page dpi, re-render that page at

higher dpi and crop, rather than upscaling.

• Cross-check visually-read values (套数, 面积, 标高) against any text-extracted

sources before writing them into spreadsheets.