Skip to content

LOG.md

Latest commit

 

History

History
54 lines (43 loc) · 2.78 KB

LOG.md

File metadata and controls

54 lines (43 loc) · 2.78 KB
Raw

LOG

2026-06-25 — Initial DWG→multipage PNG renderer + indexer

Goal: Build a pure-Rust SaaS backend library that indexes DWG files — render every layout to PNG and extract searchable metadata — verified on real sample files with proof in an HTML report.

Motivation: Indexing DWGs needs both per-sheet previews and searchable fields (text, title-block attributes, layers, layouts). Chose acadrust (MPL-2.0, pure Rust) over LibreDWG to avoid GPL + C FFI + the crashy C parser.

Changes:

  • acadrust 0.4.0 adapter (lib.rs): failsafe DWG read, catch_unwind around parse and render for crash isolation.
  • Pure geometry layer: entity → render primitives with arc/circle/ellipse/bulge sampling, INSERT block recursion (composed affine + ByBlock color), TrueType vector text (ab_glyph).
  • Page model: Layout objects → Model space + paper sheets. Paper-space viewport projection of model geometry (position/scale/twist + clip mask).
  • tiny-skia CPU renderer: world→screen transform, color-batched stroke/fill, hairline widths, Y-flip, viewport clip masks.
  • metadata.rs → IndexDoc JSON (histogram, layers, blocks, attributes, text, layouts). report.rs → self-contained HTML with thumbnails + PDF comparison.

Verification status: PASS. 11/11 real DWG 2018 files (4–46 MB) parsed, 23 pages rendered, 0 crashes. Corrupt/recovered file degraded via failsafe. Side-by-side vs plotted PDF matches network topology and building footprints. Metadata extracts real cadastral attributes and Cyrillic text correctly.

Open / next: Hatch fills, dimension geometry, leaders, lineweights, per-viewport UCS/layer-freeze, De Boor spline evaluation, text alignment. Productionize acadrust risk: vendor/fork, corpus regression, sandboxed worker.

2026-06-25 — Label readability for VLM consumption

Goal: Make every label readable by a VLM / understand the blueprint 100%.

Motivation: Tiny CAD labels are sub-pixel against a large extent, and VLMs downscale on input, so raster resolution alone can't make all labels legible.

Changes:

  • Text layer: capture every text run (Text, MText, Attribute, Dimension, MultiLeader) as a Label with world placement, project it through viewports, and emit labels/*.json with exact strings + pixel x,y,w,h, rotation, color, layer. Lossless (no OCR).
  • Resolution: --size (default 3000, up to 12000).
  • Tiling: --tiles slices each page into ~1500 px crops (tiny-skia clone_rect) with per-tile labels — readable VLM input.
  • Report: toggleable inline-SVG text overlay (crisp at any zoom) + tile grids.

Verification: PASS. Full batch at --size 5000 --tiles: 11/11 files, 23 pages, 216 tiles, 128,410 position-referenced labels (pipe-spec labels legible in tiles). 8 unit tests pass, 0 warnings.