docs/ARCHITECTURE.md

Architecture — DDI Checker

Technical reference index. Each topic is in its own document:

Document	Contents
pipeline.md	Step-by-step pipeline (parse → graph → train → serve), run commands, file structure
model_architecture.md	GNN architecture (HeteroGraphSAGE + NCN), node features (980-dim), PyG data objects, ablation results
api_reference.md	All REST endpoints, request/response formats, source field values
data_schema.md	27 normalised tables from DrugBank XML, key statistics, FK conventions
responsible_ml.md	RM1 explainability · RM2 fairness · RM3 privacy · RM4 robustness

For operational setup (Docker, env vars, quick start) see the main README.

---

Technology Stack

Backend

Library	Version	Role
Python	3.10	Runtime
Flask	2.x	REST API + Jinja2 server-side rendering
PyTorch	2.2	GNN inference + tensor ops
PyTorch Geometric (PyG)	2.x	Heterogeneous graph data structures + SAGEConv
scikit-learn	1.x	Logistic Regression baseline, StandardScaler
sentence-transformers	2.x	PubMedBERT embeddings (S-PubMedBert-MS-MARCO, 768-dim)
FAISS (faiss-cpu)	1.7	Vector similarity search — 824K DDI embeddings, on-demand
Groq Python SDK	0.9+	LLM calls — NER extraction + plain-language explanations
Pandas / NumPy	2.x / 1.x	Feature engineering, data pipeline, evaluation

Frontend

Technology	Role
Vanilla HTML / CSS / JS	Zero framework — no build step required
Jinja2 templates	Server-side rendering, auto-escaped output (XSS safe by default)
CSS Custom Properties + Grid/Flexbox	Consistent design system, glassmorphism cards
Native Fetch API	Async drug autocomplete, live dashboard polling (/api/stats)

Infrastructure

Tool	Role
Docker	Containerised deployment — Dockerfile in repo root
HuggingFace Spaces	Live demo hosting (https://marwadeeb-ddi-checker.hf.space)
Git LFS	Large binary files: *.pt model files (2–46 MB each)
GitHub	Source control