HVAC-BENCH · VERIFIED CORPUS
We don't ask you to take our word for it.
HVAC-Bench measures the gap between a generalist LLM and an LLM connected to the AskMarcel corpus. Numbers, methodology, manufacturer sources — not marketing promises.
HVAC-BENCH
BARE LLM VS ASKMARCEL
Accuracy score (%) per real HVAC scenario
| Scenario | LLM only | With AskMarcel | Gap |
|---|---|---|---|
| Code erreur exact (lookup) | 34% | 97% | +63 pts |
| Diagnostic multi-étapes | 22% | 91% | +69 pts |
| Référence manuel (page/section) | 8% | 94% | +86 pts |
| Marque hors training LLM | 41% | 89% | +48 pts |
| Hallucination critique évitée | 52% | 96% | +44 pts |
THE CORPUS
400,000+ CODES. MANUFACTURER-SOURCED.
A living pipeline: crawl, human curation, quality gates.
0
brands
0
models
0+
error codes
~0 ms
avg. latency
Crawl
Continuous collection of public and technical manufacturer documentation.
Curation
Human structuring: brand, model, code, procedure, manual reference.
Quality gates
Verification before publication — no generated content without a source.
DETERMINISM
ON A HIGH-RISK DOMAIN, WE DON'T GUESS.
The get_error_code tool performs an exact lookup in the corpus — not free-form generation. Every answer returns a manual reference with page and section. That's the difference between a manufacturer-sourced answer and a plausible hallucination.
- Exact lookup by brand, model and error code
- Manual reference + page + section on every answer
- No embedded render surface — your stack displays, we source
POST /tools/get_error_code
{
"brand": "Daikin",
"model": "Altherma 3",
"error_code": "7H"
}
→ lookup exact (pas de génération)
→ {
"diagnostic": "...",
"source": {
"manual": "Manuel installateur",
"page": 42,
"section": "4.1"
}
}