Skip to content

Grafix

Grafix is a web-based editor for business diagrams with chat input. The application translates natural-language descriptions into ready-to-use diagrams, combining LLM-based intent recognition with a rule-based template engine. Layout generation is deterministic, so identical inputs yield identical results. A session history keeps every step traceable and reversible.

At a glance

  • Describe diagrams in chat and receive an editable graphic immediately, without operating drawing tools.
  • Choose from 27 diagram types — from flowchart and SWOT to organisation chart, Venn diagram, timeline, and KPI dashboard.
  • Modify individual regions of a diagram in a targeted way, for example only the strengths list in a SWOT analysis.
  • Edit existing diagrams directly on the canvas or fine-tune them through form fields.
  • Inspect the full history of a session, jump back to earlier states, and export the session as JSON.
  • Save finished diagrams as PNG for distribution or as JSON for later editing.

Highlights

Grafix differs from a direct LLM prompt or a simple script in that the diagram is not drawn by the language model but by a template engine. The language model is used exclusively to interpret the request. This separation has direct consequences for the quality and reliability of the results.

  • Deterministic layout generation — Positions, spacing, and proportions are computed by the template engine, not suggested by the language model. Identical inputs result in identical diagrams.
  • Multi-agent pipeline — A request passes through intent recognition, execution, validation, and consistency check as separate stages. Only a single LLM call is issued per request; all other stages run rule-based.
  • Structured intent recognition — The language model classifies a request into predefined action types (create, modify slot, modify style, add element, switch template, and others) and returns the parameters as JSON. Low-confidence cases trigger clarification questions instead of hallucinations.
  • Broad set of templates — 27 diagram types in the categories process, list, hierarchy, relationship, matrix, pyramid, and cycle cover the typical use cases in teaching, administration, and project work.
  • Slot-based editing — Each template defines named regions (slots) that can be addressed individually. A change to a SWOT quadrant list affects only that quadrant and leaves the rest of the structure untouched.
  • Template upgrades — When the content of a request exceeds the capacity of the current template, the pipeline automatically switches to a more suitable variant (for example from a two-level to a three-level organisation chart) and transfers the existing content.
  • Session history as audit trail and context — Every step is documented as JSON. The history serves both as an audit trail and as an implicit few-shot context for subsequent requests, so that references such as "as before, but with four elements" can be resolved.
  • Template-aware consistency check — A downstream check identifies faulty layouts and corrects them. For structured templates, automatic interventions are suppressed in order not to alter the template's geometry.
  • Connection to an internal LLM endpoint — The application talks to an internally operated LLM service through an OpenAI-compatible interface. Data does not leave the operating environment.