Files
Nikolay MiroshnichenkoandClaude Opus 4.8 2d0a63c548 docs: add spec-PDF→KNX case study (EN+RU, anonymised)
A 96% structural match of an auto-designed GA structure vs a professional
reference, from a 42-page spec alone — 662 GA, 0 validation errors, HA package.
No client data/files; reference used only as an anonymous benchmark.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-30 23:27:32 +02:00

2.8 KiB

Case study — from a spec PDF to a validated 662-GA KNX structure, in minutes

🇷🇺 Русская версия: case-study.ru.md

The question

How far can a project specification alone be turned into a finished, standards-correct KNX group-address structure — and how close would it get to professional work?

Input

A real residential project — 14 rooms (~155 m²), Zennio-based: switched + dimmed lighting, electric underfloor heating, ventilation with heat recovery, motorised curtains, water / electricity / heat metering, leak & fire safety — described only in a 42-page technical specification (PDF). No group addresses, no DPTs, no naming were given: the spec lists equipment and functions, not a KNX structure.

The workflow

spec PDF
   │  an AI assistant, grounded in the KNX Association standard + industry practice,
   ▼  DESIGNS the full group-address structure
GA structure (taxonomy · naming · DPTs · command/status pairs · scenes · reserves)
   │  nickol-knx-mcp VALIDATES it and GENERATES the outputs
   ▼
Home Assistant package  +  ETS-importable group-address XML

Output

  • 662 group addresses across 10 functional domains, as a ga-export/01 XML ready for ETS Import Group Addresses → place devices → export .knxproj.
  • Produced in minutes.

The check — against a professional reference implementation of the same project

Metric Result
Structural match 96 % (662 vs 687 GA)
Domain taxonomy 10 / 10 identical
Address names 118 byte-identical, 68 % with a close analog
DPT discipline exact sub-types throughout; DPT 5.001 ≈ 1 : 1
Validation (nickol-knx-mcp) 0 errors; reserve / logic noise auto-classified
Home Assistant 9 climate zones · 11 shutters · 12 dimmers + sensors, generated directly

What it shows

The standardised, time-consuming "skeleton" of a KNX project — the part defined by the standard and the equipment — can be auto-designed to near-professional quality in minutes, vs the days of manual group-address work. That frees the engineer's time for the genuinely creative part: the project-specific logic, which stays the engineer's craft (the residual few % are exactly those one-off logic functions).

Honesty notes

  • No client data or project files are shared. The reference implementation is used only as an anonymous benchmark; the case study reports parameters and match figures, nothing identifying.
  • The design is produced by an AI assistant grounded in KNX knowledge (the KNX Association standard + broad industry practice). nickol-knx-mcp's role is read-only validation and Home Assistant / ETS export — it never connects to a live bus.
  • Figures are from a structure that actually passed the tool's own checks (0 errors).