OQVAR — ask the drawing, check the answer

What it reads / four kinds of drawing

One question. Every kind of drawing.

CAD plans, BIM models, scanned sheets and PDF schedules. Each question goes to the engine that can answer it exactly, and each answer points back to the drawing.

Drawing types 04 sources
A-102.dxfIllustration / 01
LIVING32.4 m²KITCHEN18.6 m²BEDROOM 114.8 m²BATH6.1 m²SAUNA4.2 m²BEDROOM 211.9 m²HALL9.7 m²18.6 m²7 4006 800
18.6 m²Kitchen · from the room polygon
Computed, not estimated↗ Geometry
CAD · DXFSwiss dwellings

Floor plans

Rooms, doors, windows and fixtures are read from the drawing’s own geometry. Areas come from the room polygon, and every counted item is outlined where it is on the sheet.

  • Geometry engine
  • 0 LLM tokens
  • ~8 ms per question
  • EN · DE · FR · IT
Measured

99.49% of 37,584 questions answered correctly.

A real question to tryHow many windows are drawn on this floor plan?
Explore

Previews are illustrations. The questions, the accuracy figures and the drawings behind each Explore button come from Stoa’s dataset.

Why OQVAR / compute, don’t guess

An answer that only sounds right is useless on site.

Site engineers spend hours a day searching dense drawing sets. General-purpose AI reads those drawings as pictures, and it shows. Same sheet, two answers:

General-purpose AI

Illustration

“There are about 9 doors on this floor.”

  • Miscounts symbols. Doors and fixtures are estimated from pixels.
  • Misreads scales. An area is only right if the scale is.
  • Invents dimensions. A plausible number with nothing behind it.
Nothing on the sheet to check

OQVAR

Same sheet
7doorsGeometry · 0 LLM tokens · 8 ms
LIVING32.4 m²KITCHEN18.6 m²BEDROOM 114.8 m²BATH6.1 m²SAUNA4.2 m²BEDROOM 211.9 m²HALL9.7 m²12345677 4006 800
Every door outlined and numbered on the sheet
How it answers / cheapest trustworthy engine first

Exact when it can be. Honest when it can’t.

Each question goes to the first engine that can answer it exactly. A language model steps in only where the drawing has no structure to compute from, and the answer always says which engine answered.

Engines 05 in order
  1. 1A

    Geometry / BIM

    Dataset-style questions on CAD and IFC drawings

    No LLM

    Polygons on the sheet, IFC elements in 3D

  2. 1B

    Detection hybrid

    Counts on raster plans

    No LLM at question time

    The network’s detected shapes

  3. 2

    Query planner

    Free-form questions on CAD and IFC

    1 query · ~1,200 tokens

    The rows the query returned

  4. 3

    Sheet text

    PDF sheets with real text, such as schedules

    Picks rows, we count

    Each picked row, highlighted

  5. 4

    Vision

    Scans and anything without structure

    Reads the image

    The model’s own boxes, flagged low confidence

Benchmark / Stoa’s own questions

Measured, not claimed.

Counts must be exact; areas, lengths and volumes within 2%. The weak spots are on the chart too: scanned plans and real PDF schedules are where the work continues.

Share of questions answered correctlyby drawing type

Every number, every baseline model and how each was measured: open the app and press Test report.

Ready when you are / no sign-up

Open a drawing. Ask it anything.

The drawing register, the linked 3D model and the full test report are one click away.

Explore