Skip to content

Diagramm-Generator

Diagramm-Generator is a browser-based application for creating professional diagrams from natural-language descriptions and structured data. The application combines a multi-stage agentic pipeline (intent recognition, planning, execution) with a multi-agent LLM system, deterministic code generation, and a validation and correction mechanism. Output is produced as Mermaid, draw.io, or Gantt diagrams with export to SVG, PNG, or source code.

At a glance

  • Generate diagrams from free-form descriptions — a description in natural language is sufficient to obtain a suitable diagram.
  • Visualize existing data — CSV, JSON, Markdown, or indented lists are processed without manual preparation.
  • Refine diagrams iteratively — structures, labels, and connections can be adjusted step by step through the chat.
  • Start from prefabricated templates — a library of templates for business, technical, and project contexts accelerates onboarding.
  • Reuse results — export as SVG, PNG, or Mermaid source code for inclusion in documentation, presentations, and wikis.
  • Compare multiple diagram variants — gallery and history allow comparison of different suggestions and revisions.
  • Save and resume sessions — work states can be exported, imported, and continued elsewhere.

Highlights

In contrast to a direct LLM prompt or a simple script-based generator, the application structures the generation process into clearly separated stages, automatically checks the result, and corrects it where necessary. This produces reproducible and syntactically correct diagrams even from ambiguous or incomplete input.

  • Agentic IPE pipeline — A three-stage process with intent recognition, planning, and execution separates the understanding of the request, the structural planning, and the code generation. This allows error sources to be localized and individual stages to be improved independently.
  • Multi-agent LLM system — Chat, diagram, and validation agents each use their own LLM configuration with adapted parameters (temperature, token limits, sampling). Conversational tasks receive softer settings; code generation and validation receive more deterministic ones.
  • Deterministic code generation — Mermaid code is produced from the structured plan via a rule-based converter, without the LLM writing the code directly. This improves reproducibility and syntactic fidelity.
  • Validation and correction loop — Generated code is validated before rendering. On syntax errors, a validation agent corrects the code with reference to a complete Mermaid syntax reference; on repeated failure, a template fallback applies.
  • Domain defaults for typical processes — Common lanes and activities are pre-defined for IT support, HR, and sales. Minimal input such as "support process with 1st and 2nd level" produces a fully populated draft.
  • Smart Convert across multiple input formats — Uploaded or entered data in CSV, JSON, Markdown lists, or indented hierarchies is automatically detected and transformed into a suitable diagram representation.
  • Connection to three diagram engines — Mermaid (with CLI renderer), draw.io (via N2G), and python-gantt are addressed through a common interface. A renderer factory selects the appropriate engine per diagram type.
  • Broad coverage of diagram types — Flowchart, mindmap, org chart, sequence, Gantt, swimlane, kanban, state diagram, ER diagram, and network graph are supported equally.
  • Configurable processing — Parameters such as maximum correction attempts, JSON intermediate format for structured diagram types, validation depth, and domain defaults are controlled via a YAML configuration.