Architectural Drawing Parser

Overview

Vision AI pipeline to extract structured building data from architectural drawings, floor plans, and IBC/IRC code compliance documents. Uses Claude's vision capabilities to read and interpret professional drawings, returning a normalized JSON object suitable for downstream 3D modeling or code validation workflows.

Supports IBC occupancy types (A-1 through U), construction types (I-A through V-B), sprinkler systems (NFPA 13/13R/13D), building dimensions, unit breakdowns, egress data, and floor plan elements (rooms, walls, doors, windows).

Instructions

Supported Drawing Types

Output Data Structure

The parser returns a BuildingData JSON object with these fields:

• occupancy -- IBC occupancy type (e.g., "R-2", "A-2", "B")
• constructionType -- IBC construction type (e.g., "V-B", "I-A")
• sprinklerSystem -- "NFPA 13", "NFPA 13R", "NFPA 13D", or "None"
• stories -- { permitted, actual }
• height -- { permitted: { feet, meters }, actual: { feet, meters } }
• totalBuildingArea -- { sqft, sqm }
• units -- Array of { name, area: { sqft, sqm }, occupantLoad, loadFactor, count }
• travelDistances -- Array of { floor, maximum: { feet, meters } }
• scale -- Scale notation string (e.g., 1/16" = 1'-0")
• rooms -- Array of { name, type, estimatedArea, dimensions } (floor plans only)

Parsing Approach

1. Send the drawing to Claude's vision API with a structured extraction prompt: put the image (or PDF) block first and the instructions after it
2. Request all building data as a single JSON object and state the schema in the prompt
3. Convert all areas to both sqft and sqm (1 sqft = 0.0929 sqm)
4. Convert all distances to both feet and meters (1 foot = 0.3048 m)
5. Parse the JSON from the response text and reject the result if required fields are missing

import anthropic, base64, json, pathlib

client = anthropic.Anthropic()   # reads ANTHROPIC_API_KEY from the environment
page = base64.standard_b64encode(pathlib.Path("A-101-level-1.jpg").read_bytes()).decode()

message = client.messages.create(
    model="claude-opus-5-5",     # any current vision-capable Claude model
    max_tokens=4096,
    messages=[{"role": "user", "content": [
        {"type": "image", "source": {"type": "base64", "media_type": "image/jpeg", "data": page}},
        {"type": "text", "text": "Extract occupancy, constructionType, sprinklerSystem, stories, height, "
            "totalBuildingArea, units, travelDistances, scale and rooms from this drawing. Give areas in sqft and sqm, "
            "distances in feet and meters. Use null for anything not shown. Reply with one JSON object only."},
    ]}],
)
building = json.loads(message.content[0].text)

For a whole PDF set, send it as a document block (base64 or a Files API file_id) instead of an image; the API takes up to 32 MB and 100 pages per request on 200k-context models (600 on 1M-context models). Per image, the limits are 10 MB and 8000x8000 px, and images are downscaled to a 1568 px long edge on standard models (2576 px on Claude 4.7 and later), so tiny dimension text on a large sheet can be lost.

Best Practices

• Use 150 DPI or higher for scanned drawings; keep text legible after the long edge is capped
• JPEG or PNG; for sheets with dense notes convert PDFs to images and crop to the table or plan area (pdftoppm -png -r 200 -f 3 -l 3 drawing.pdf sheet)
• Process multi-sheet PDFs one page at a time, then merge results
• Label each image ("Sheet A-101:") when sending several in one request
• Treat dimensions read from linework as approximate: prefer dimensions printed on the sheet over measuring from the scale
• Always verify extracted data against the source before structural calculations

Examples

Example 1: Parsing a Floor Plan PDF

A developer receives a scanned floor plan of a 2-bedroom apartment unit and needs room dimensions for a renovation estimate.

Input: apartment_unit_plan.jpg (scanned at 200 DPI, 1/4" = 1'-0" scale)

Extracted JSON:
{
  "rooms": [
    { "name": "Living Room", "type": "living", "estimatedArea": { "sqft": 240, "sqm": 22.3 }, "dimensions": { "width": 16, "depth": 15, "units": "feet" } },
    { "name": "Kitchen", "type": "kitchen", "estimatedArea": { "sqft": 120, "sqm": 11.1 }, "dimensions": { "width": 12, "depth": 10, "units": "feet" } },
    { "name": "Master Bedroom", "type": "bedroom", "estimatedArea": { "sqft": 168, "sqm": 15.6 }, "dimensions": { "width": 14, "depth": 12, "units": "feet" } },
    { "name": "Bedroom 2", "type": "bedroom", "estimatedArea": { "sqft": 132, "sqm": 12.3 }, "dimensions": { "width": 12, "depth": 11, "units": "feet" } },
    { "name": "Bathroom", "type": "bathroom", "estimatedArea": { "sqft": 48, "sqm": 4.5 }, "dimensions": { "width": 8, "depth": 6, "units": "feet" } }
  ],
  "scale": "1/4\" = 1'-0\""
}

The developer uses the room dimensions to calculate material quantities for flooring (708 sqft total) and wall paint coverage.

Example 2: Extracting Building Data from an IBC Compliance Drawing

An architect submits a code compliance sheet for a 3-story apartment building. The parser extracts all building classification and egress data.

Input: ibc_compliance_sheet.jpg (building area analysis table + egress diagram)

Extracted JSON:
{
  "occupancy": "R-2",
  "constructionType": "V-B",
  "sprinklerSystem": "NFPA 13",
  "stories": { "permitted": 4, "actual": 3 },
  "height": {
    "permitted": { "feet": 60, "meters": 18.29 },
    "actual": { "feet": 35, "meters": 10.67 }
  },
  "totalBuildingArea": { "sqft": 8910, "sqm": 827.9 },
  "units": [
    { "name": "Type A", "area": { "sqft": 834, "sqm": 77.5 }, "occupantLoad": 5, "loadFactor": "1/200 SF", "count": 6 },
    { "name": "Type B", "area": { "sqft": 645, "sqm": 59.9 }, "occupantLoad": 4, "loadFactor": "1/200 SF", "count": 6 }
  ],
  "travelDistances": [
    { "floor": "Level 1", "maximum": { "feet": 66, "meters": 20.1 } },
    { "floor": "Level 2", "maximum": { "feet": 66, "meters": 20.1 } },
    { "floor": "Level 3", "maximum": { "feet": 66, "meters": 20.1 } }
  ]
}

This data feeds into the ibc-building-codes skill for compliance validation and the spec-to-3d skill for 3D model generation.

Guidelines

• Accuracy depends on drawing quality and image resolution; low-res scans may produce incorrect dimensions
• Very small text (title blocks, fine notes) may be misread -- zoom in for detail drawings
• Complex overlapping hatching or linework may confuse room detection
• Proprietary symbols or non-standard abbreviations may not be recognized
• Always treat extracted data as an estimate; verify critical measurements manually
• For multi-sheet sets, parse each sheet separately and merge the structured data
• Permitted stories, heights and areas depend on the IBC edition and local amendments; read them from the sheet rather than inferring them from the occupancy and construction type
• The parser works best with US-standard architectural drawings; metric-only drawings may need prompt adjustments