diagram-analyser

Structural analysis of diagrams — the lens-family member that reads diagrams as graphs (nodes + edges) rather than as raster images.

Supports four text formats out of the box:

Plus an optional [vision] path for image diagrams (.png/.jpg) via Anthropic Claude Vision — explicit-only, since those extensions auto-route to image-analyser.

image-analyser sees raster pixels; this one parses the diagram's structure — nodes, edges, orphans, cycles, depth, naming. Built for design-unit assessment (CS/IS architecture, ER, UML, flowcharts).

Install

pip install diagram-analyser                       # text formats — pure Python
pip install 'diagram-analyser[vision]'             # adds image-diagram support

For the vision path, set one of:

export ANTHROPIC_API_KEY=sk-ant-...

Use

Python:

from diagram_analyser import DiagramAnalyser

result = DiagramAnalyser().analyse("architecture.mmd")
print(result.diagram_type)              # "flowchart"
print(result.graph.node_count)          # 7
print(result.structure.orphan_nodes)    # []
print(result.structure.cycle_count)     # 1
print(result.structure.is_dag)          # False

# Image diagram (needs [vision] extra + ANTHROPIC_API_KEY):
result = DiagramAnalyser().analyse("diagram.png")
print(result.vision.raw_description)

CLI:

diagram-analyser flow.mmd                  # human summary
diagram-analyser model.puml --json         # raw JSON
diagram-analyser arch.png                  # vision path (needs API key)
diagram-analyser serve                     # HTTP API on port 8013
diagram-analyser manifest                  # capability manifest

HTTP (diagram-analyser serve on port 8013):

curl -F file=@flow.mmd http://localhost:8013/analyse
curl http://localhost:8013/health

Signals

For every diagram (text or vision-extracted):

• Graph — nodes (id, label, kind), edges (source, target, kind, label), counts.
• Structure — orphan_nodes (disconnected), cycle_count + sample cycles, max_depth, is_dag, connected_components.
• Naming quality — label coverage, average label length, suspiciously short labels.
• Diagram type — for mermaid: flowchart / classDiagram / erDiagram / sequenceDiagram / stateDiagram; for PlantUML: uml / mindmap / wbs; for graphviz: digraph / graph; for drawio: best-effort from cell hints.

The family

Part of the lens analyser family.

Limits

• v1 text parsers are pragmatic — not full grammars. They get nodes, edges, and the diagram type for ~all realistic student diagrams; corner cases (subgraphs, deeply nested mermaid classes) may parse approximately.
• Vision path produces structure only as well as the LLM extracts it. We surface the raw description so you can verify.

License

MIT