---
title: Diagram Structure Extractor
emoji: 📊
colorFrom: blue
colorTo: green
sdk: docker
app_port: 8000
pinned: false
license: mit
---

# Diagram Structure Extractor

Upload an architecture diagram and get its structure back as schema-valid JSON:
every text label, box, arrow, icon, and the relationships between components.
Output is a typed Pydantic model, so the JSON parses on every call.

**Demo**: the Space opens on a Streamlit app. Upload a diagram or pick one
of the samples to see the annotated overlay, the relationship graph and the
JSON, with JSON and CSV downloads. The sidebar switches each detector stage.

**API** (same host)

- `POST /extract` — multipart image upload; returns the typed structure, an
  annotated PNG, and a flat edge CSV
- `GET /health` — liveness
- `GET /docs` — interactive API docs

**Measured results** (relationship F1; edges are directed, and a double-headed
connector counts once each way)

- 14 synthetic reference architectures: **1.000**
- A real draw.io export: **0.900**
- 10 held-out draw.io-style diagrams: **0.789** (0.711 when first scored,
  before the code-review fixes)
- Schema-valid JSON on every input, at every pipeline stage
- Detectors: PaddleOCR text (F1 0.962), Canny boxes (1.000), directed-lines
  arrows (1.000 on the synthetic set), template-matched icons (1.000)

Handles draw.io conventions: thin dashed connectors, small and double-headed
arrowheads, elbow routing, line labels, icon nodes with captions underneath.

Source: https://github.com/Mark007-R/Diagram-Structure-Extractor