CoDesign

CoDesign is a research prototype for AI-assisted P&ID knowledge graph exploration and design-decision traceability. The project converts DEXPI-based P&ID examples into structured data, lets users ask graph-grounded questions through GraphRAG, and records human or AI-agent design decisions as reviewable logs and candidate graph updates.

The goal is not to automatically produce final approved engineering P&IDs. The goal is to make P&ID knowledge, assumptions, review gates, and human/AI decision traces explicit and reusable.

What The App Can Do

• Select a DEXPI-derived P&ID example from the dataset.
• Visualize the selected P&ID as a node-link graph.
• Ask GraphRAG questions such as what is connected to P-4713?.
• Run a five-question design decision workflow.
• Separate human-confirmed responses from AI-default assumptions.
• Update the candidate graph as decisions are recorded.
• Export decision logs as CSV.
• Export decision-generated candidate graph data as JSON.

Main Workflow

1. Choose an input P&ID from the left panel.
2. Use GraphRAG Mode to ask questions about equipment, lines, tags, nozzles, and connectivity.
3. Start the design assistant conversation.
4. Answer five design questions:
   • fluid/material basis,
   • design flow rate,
   • pump duty and pressure protection,
   • treatment configuration,
   • human review gate.
5. Review how the graph changes as decisions are recorded.
6. Export the decision log or candidate graph dataset.

Repository Structure

codesign/
├── data/
│   ├── pids/                 # normalized DEXPI/Proteus XML P&ID inputs
│   ├── json/                 # pyDEXPI JSON model exports
│   ├── gexf/                 # graph exports for graph analysis
│   ├── schemas/              # schemas for graph, decision log, and candidate exports
│   ├── design_decisions/     # sample human and AI-agent decision logs
│   └── candidates/           # sample decision-generated candidate graph exports
├── public/sample/            # default PFAS treatment example used by the app
├── scripts/                  # conversion and database-building scripts
├── server/                   # Express API for dataset loading, GraphRAG, and agent generation
├── src/                      # React/Vite frontend
├── .env.example              # environment variable template
├── package.json
└── README.md

Data Included In The Repository

The repository includes a working sample data package for development and review:

• data/pids/: 35 normalized XML P&ID inputs.
• data/json/: 35 pyDEXPI JSON exports.
• data/gexf/: 30 GEXF graph exports. Some are sparse or empty because the current graph export pipeline does not fully capture every DEXPI example structure.
• data/schemas/: lightweight schemas for the exported data formats.
• data/design_decisions/: sample human and AI-agent decision logs.
• data/candidates/: sample candidate graph exports.

The pyDEXPI JSON files are the primary derived data. GEXF files are secondary graph-analysis exports.

API Endpoints

The local API server provides:

• GET /api/designs: list available input P&ID JSON files.
• GET /api/designs/:id: load one pyDEXPI JSON model.
• POST /api/graphrag/query: answer a graph-grounded question using retrieved graph evidence.
• POST /api/agent/generate-decision-log: generate sample AI-agent decision-log entries for review.
• GET /api/health: check API status and whether an OpenAI key is configured.

Setup

Install dependencies:

npm install

Create a local .env file from the template:

cp .env.example .env

Add your API key if you want LLM-backed GraphRAG and agent generation:

OPENAI_API_KEY=your_key_here
OPENAI_MODEL=gpt-5-mini
API_PORT=8787
VITE_TOKEN_BUDGET=20000
VITE_OPENAI_MODEL=gpt-5-mini

Run the API and frontend together:

npm run dev:full

Or run them separately:

npm run api
npm run dev -- --host 127.0.0.1

The frontend usually opens at http://127.0.0.1:5173. The API runs at http://127.0.0.1:8787.

Build

npm run build

Transformation Scripts

The conversion scripts are in scripts/.

Convert DEXPI/Proteus XML files to pyDEXPI JSON and GEXF:

python scripts/convert_dexpi_to_kg.py --input-dir data/pids --json-dir data/json --gexf-dir data/gexf

Build a SQLite-style P&ID knowledge graph database from pyDEXPI JSON:

python scripts/build_pid_kg_db.py --json-dir data/json --db-path data/pid_kg.sqlite

Research Scope

CoDesign should be described as an AI-assisted decision support and data-generation prototype. It supports:

• P&ID knowledge graph structuring,
• evidence-grounded GraphRAG questions,
• design decision logging,
• human/AI response source tracking,
• reviewable candidate graph export.

It does not produce final approved engineering P&ID drawings and does not perform standards-certified DEXPI XML write-back.

Source Data

The upstream source dataset is the public DEXPI repository Public Example PIDs / TrainingTestCases:

https://gitlab.com/dexpi/TrainingTestCases

Most normalized XML inputs in this project were prepared from the DEXPI 1.3 example PIDs:

https://gitlab.com/dexpi/TrainingTestCases/-/tree/master/dexpi%201.3/example%20pids?ref_type=heads

This project does not claim authorship of the original P&ID examples. CoDesign creates derived research artifacts from those public examples, including normalized XML inputs, pyDEXPI JSON exports, GEXF graph exports, SQLite tables, decision logs, and candidate graph datasets.