Home Plan Compliance Checker — 3-Agent Pipeline

A fully agentic system to read architectural home plans (PDFs or images), extract structured data, and validate them against predefined building code rules.

---

Architecture

┌─────────────────────────────────────────────────────────────────┐
│  INPUT: Plan PDF/Image  +  Rules Text File                      │
└───────────────┬─────────────────────────────────────────────────┘
                │
                ▼
┌──────────────────────────────────────────┐
│  AGENT 1: GPT-4o Vision Extractor        │
│  • Renders each PDF sheet at 300 DPI     │
│  • Sends image to GPT-4o (high detail)   │
│  • Extracts: rooms, doors, windows,      │
│    stairs, setbacks, annotations         │
│  • Output: structured JSON per element   │
│  • Flags low-confidence extractions      │
└───────────────┬──────────────────────────┘
                │  extracted_plan.json
                ▼
┌──────────────────────────────────────────┐
│  AGENT 2: Vectorization Engine           │
│  • Converts JSON elements → text chunks  │
│  • Parses rules text → individual rules  │
│  • Embeds both via text-embedding-3-large│
│  • Builds two FAISS indexes              │
│    (L2-normalized for cosine similarity) │
│  • Saves indexes + chunks to disk        │
└───────────────┬──────────────────────────┘
                │  FAISS indexes + chunk metadata
                ▼
┌──────────────────────────────────────────┐
│  AGENT 3: Compliance Reasoning Engine    │
│  For each rule:                          │
│   1. Embed rule → retrieve top-K plan    │
│      chunks via cosine similarity        │
│   2. Pass rule + chunks to GPT-4o for    │
│      logical compliance determination    │
│   3. If INFO_NOT_FOUND → feedback loop   │
│      re-queries Agent 1 with targeted    │
│      question about missing element      │
│  Output: COMPLIES / NON_COMPLIANT /      │
│          INFO_NOT_FOUND per rule         │
└───────────────┬──────────────────────────┘
                │
                ▼
┌─────────────────────────────────────────────────────────────────┐
│  OUTPUT: compliance_summary.json  +  detailed_results.json      │
└─────────────────────────────────────────────────────────────────┘

---

Installation

pip install -r requirements.txt

Set your OpenAI API key:

export OPENAI_API_KEY=sk-...

---

Usage

Run full pipeline (recommended)

python run_pipeline.py --plan /path/to/homeplan.pdf --rules sample_rules.txt

Run agents individually

Agent 1 only (extract plan data):

python agent1_extractor.py /path/to/homeplan.pdf
# Output: extracted_plan.json

Agent 2 only (vectorize, requires Agent 1 output):

python agent2_vectorizer.py extracted_plan.json sample_rules.txt
# Output: vector_store/

Agent 3 only (compliance check, requires Agent 2 output):

python agent3_compliance.py vector_store/ /path/to/homeplan.pdf
# Output: compliance_results/

Disable the re-query feedback loop

python run_pipeline.py --plan plan.pdf --rules rules.txt --no-requery

---

Output Files

File	Description
output/extracted_plan.json	Structured extraction from Agent 1
output/vector_store/	FAISS indexes and chunk pickles from Agent 2
output/compliance_results/compliance_summary.json	High-level report: overall verdict, violations, gaps
output/compliance_results/detailed_results.json	Per-rule breakdown with reasoning and evidence

---

Verdict Meanings

Verdict	Meaning
COMPLIES	Plan data meets the rule requirement
NON_COMPLIANT	Plan data clearly violates the rule
INFO_NOT_FOUND	Relevant data absent from plan (manual review needed)

Overall plan verdicts:

• COMPLIES — All rules pass
• NON_COMPLIANT — At least one rule is violated
• PARTIAL_REVIEW_NEEDED — No violations found but some rules couldn't be evaluated due to missing data

---

Rules File Format

Rules can be in either format:

Numbered rules:

1. Minimum bedroom dimension shall be 7 feet in any direction.

2. Maximum riser height is 7.75 inches.

Free text blocks (separated by blank lines):

All egress windows shall have a minimum net clear opening of 5.7 square feet.

Stairway width shall not be less than 36 inches.

---

Key Design Decisions

• GPT-4o high detail mode is used for all plan images — critical for small annotations and rotated dimension text
• FAISS cosine similarity retrieves relevant plan elements per rule (similarity threshold: 0.15)
• LLM reasoning layer in Agent 3 performs actual logic/arithmetic, not just similarity matching
• Confidence flags on extracted elements are passed through to the compliance reasoning
• Feedback loop between Agent 3 and Agent 1 handles INFO_NOT_FOUND cases automatically
• Multi-sheet PDFs are fully supported — each sheet processed independently then merged