CAD Generator with RAG + LLM

Automated CAD file generation from Indonesian natural language using hybrid LLM + RAG + Regex architecture.

Author: Lis Wahyuni

Key Features

• Hybrid NLP Parser: Llama 3.2:1b (LLM) + Regex validation + ChromaDB RAG
• Multi-format Output: DXF, SVG (2D) | STL, OBJ/MTL (3D)
• Object Types: Chair, Table (round/rectangular), Sofa, Cabinet, Room, House
• Indonesian Input: Native Bahasa Indonesia with semantic translation
• Standards-based: Furniture dimensions from RAG knowledge base
• Circular Geometry: Precise circular table top & cylindrical legs

Quick Start

# 1. Install uv (fast Python package manager)
curl -LsSf https://astral.sh/uv/install.sh | sh
# 2. Install Ollama + Llama 3.2:1b
curl -fsSL https://ollama.com/install.sh | sh
ollama pull llama3.2:1b
# 3. Clone and install dependencies
git clone https://github.com/liswahyuni/CADpython.git
cd CADpython
uv sync
# 4. Run demo (generates 35 files: 7 objects × 5 formats)
uv run demo.py
# 5. Generate 3D preview GIFs (optional - converts STL to rotating GIF animations)
uv run generate_3d_gifs.py

Output: All files saved to demo_output/ directory

Note: GIF animations in examples below are generated from STL 3D files using matplotlib rendering.

Architecture

Pipeline Flow

┌─────────────────────────────────────────────────────────────────┐
│ INPUT: Indonesian Natural Language │
│ "Kursi 4 kaki, dudukan 40x40cm, tinggi 45cm" │
└───────────────────────┬─────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ STAGE 1: LLM Parser (llm_parser.py) │
│ • Llama 3.2:1b extracts structured JSON │
│ • Output: {type, dimensions, features, confidence} │
└───────────────────────┬─────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ STAGE 2: Regex Validator (llm_parser.py) │
│ • Override LLM with explicit dimensions from text │
│ • Ensures 100% accuracy for numeric values │
└───────────────────────┬─────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ STAGE 3: ChromaDB RAG (pdf_rag.py) │
│ • Semantic search: Indonesian → English standards │
│ • Retrieve missing dimensions from knowledge base │
└───────────────────────┬─────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ STAGE 4: 2D CAD Generation (cad_generator.py) │
│ • Top view + Front view with dimensions │
│ • Output: DXF (AutoCAD compatible) + SVG (web preview) │
└───────────────────────┬─────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ STAGE 5: 3D Extrusion (extruder_3d.py) │
│ • Realistic 3D mesh with features (legs, doors, windows) │
│ • Output: STL (3D printing) + OBJ/MTL (rendering) │
└─────────────────────────────────────────────────────────────────┘

Tech Stack

Component	Technology	Purpose
LLM	Llama 3.2:1b (Ollama)	Indonesian NLP + JSON extraction
Vector DB	ChromaDB + all-MiniLM-L6-v2	Semantic search (384-dim embeddings)
2D CAD	ezdxf + svgwrite	DXF (AutoCAD) + SVG generation
3D Mesh	trimesh	STL (binary) + OBJ/MTL (text)
3D Preview	matplotlib + Pillow	Rotating GIF animations

Hybrid Parser Strategy

Why Hybrid? Each component handles different aspects:

Component	Strength	Limitation
LLM	Handles ambiguity, context, inference	May hallucinate numbers
Regex	100% accurate for explicit dimensions	Cannot understand context
ChromaDB RAG	Fast retrieval of standards	Requires knowledge base

Example Processing:

Input: "Sofa 3 dudukan panjang 200cm"
│
├─ LLM: {type: "sofa", length: 2.0, seats: 1} ← Wrong seats!
├─ Regex: {seats: 3} ← Corrects "3 dudukan" → seats=3
└─ RAG: {width: 0.9, height: 0.85} ← Retrieves standards
│
Result: {type: "sofa", length: 2.0, width: 0.9, height: 0.85, seats: 3}

Key Insight: Regex overrides LLM for explicit dimensions, ensuring accuracy.

Examples

All 7 demo examples with 2D (SVG) and 3D (GIF animation) outputs:

Note: 3D GIF animations below are generated by converting STL files to rotating 360° previews using matplotlib.

1. Chair (4 Legs)

Input:"Kursi dengan 4 kaki, dudukan persegi 40x40 cm, tinggi 45 cm"
Dimensions: 0.40m × 0.40m × 0.45m

2D Output (SVG)	3D Output (GIF)

2. Room (Door & Window)

Input:"Ruangan ukuran 4x5 meter, dengan 1 pintu di sisi barat dan 1 jendela di sisi utara"
Dimensions: 4.00m × 5.00m × 3.00m

2D Output (SVG)	3D Output (GIF)

3. Round Dining Table

Input:"Meja makan lingkaran diameter 120 cm dengan 4 kaki, tinggi 75 cm"
Dimensions: Ø1.20m × 0.75m

2D Output (SVG)	3D Output (GIF)

4. Dining Table (6 People)

Input:"Meja makan untuk 6 orang"
Dimensions: 1.63m × 1.02m × 0.74m (from ChromaDB standards)

2D Output (SVG)	3D Output (GIF)

5. Three-Seat Sofa

Input:"Sofa 3 dudukan dengan sandaran dan lengan, panjang 200 cm lebar 90 cm tinggi 85 cm"
Dimensions: 2.00m × 0.90m × 0.85m

2D Output (SVG)	3D Output (GIF)

6. Wardrobe (2 Doors)

Input:"Lemari pakaian 2 pintu, ukuran 120x60 cm, tinggi 200 cm"
Dimensions: 1.20m × 0.60m × 2.00m

2D Output (SVG)	3D Output (GIF)

7. Modern House (with Garage & Windows)

Input:"Rumah modern 10x12 meter dengan garasi, 2 kamar tidur, 4 jendela, tinggi 3.5 meter"
Dimensions: 10.00m × 12.00m × 3.50m

2D Output (SVG)	3D Output (GIF)

Project Structure

CADpython/
├── demo.py # Main demo + pipeline orchestrator
├── llm_parser.py # Llama 3.2 parser
├── pdf_rag.py # ChromaDB RAG system
├── cad_generator.py # 2D DXF/SVG generation
├── extruder_3d.py # 3D STL/OBJ generation
├── generate_3d_gifs.py # STL to GIF converter
├── pyproject.toml # Dependencies
├── demo_output/ # Generated files
│ ├── *.dxf # 2D CAD files
│ ├── *.svg # 2D vector graphics
│ ├── *.stl # 3D models (binary)
│ ├── *.obj *.mtl # 3D models (text + materials)
│ └── gifs/ # 3D preview animations
│ └── *.gif # Rotating 3D previews
└── rag_documents/ # PDF knowledge base
└── chroma_db/ # ChromaDB persistent storage

References

Core Libraries:

• Ollama - Local LLM inference engine
• Llama 3.2 - 1B parameter language model
• ChromaDB - Vector database for RAG
• Sentence Transformers - Embedding models (all-MiniLM-L6-v2)

CAD & 3D Processing:

• ezdxf - DXF file generation (AutoCAD format)
• svgwrite - SVG vector graphics
• trimesh - 3D mesh processing (STL/OBJ)
• matplotlib - 3D visualization
• Pillow - Image processing for GIF generation

Supporting:

• PyPDF2 - PDF text extraction
• NumPy - Numerical computing