The right manual
H72 on an Aquarea: the tank temperature looks wrong.
SKU WH-UD09LE5 resolved: the response opens page 184 of the J-Series manual. The technician checks the sensor, connector, and PCB — not a guessed flow or gas issue.
HVAC-BENCH V1 · MEASUREMENTS FROZEN AUGUST 31, 2026
Proof before
the quote.
Put manufacturer-sourced HVAC knowledge behind your CMMS, bot, or AI agent. AskMarcel returns the applicable manual page — or abstains when the machine identity or source is missing.
AskMarcel policy
- Manufacturer excerpts only after SKU resolution
- no_exact_source and model_ambiguous are not billed
- REST, MCP, and Harness: three distinct contracts
identity resolved + exact source
→ cited procedure
missing source or ambiguous model
→ structured abstention
→ cited procedure
missing source or ambiguous model
→ structured abstention
The rule that protects your jobs
A useful system also knows when to say no.
An LLM can answer quickly and confidently while sending you to the wrong manual. AskMarcel does not sell a chattier response: it applies a provenance rule to your software.
Resolved SKU + applicable manual + covered code
→ Sourced excerpt: document, page, section, and value
Missing identity, ambiguous model, or unavailable manual
→ Structured abstention: request the missing data
The new benchmark
Proof is more than a single percentage.
18 cases were frozen before scoring. Three models, two conditions, two independent raters, and 97 scored answers. We deliberately separate retrieval from abstention.
When the exact page exists
5 documented Panasonic Aquarea codes · AskMarcel tools and usage-policy arm
| Model | Exact applicable proof | Manufacturer invention |
|---|---|---|
| GLM | 5/5 | 0/5 |
| Luna | 5/5 | 0/5 |
| Deepseek | 4/5 | 0/5 |
When the system must abstain
Missing source, brand trap, or ambiguous model · AskMarcel arm
| Model | Correct abstention | Invention |
|---|---|---|
| GLM | 8/9 | 1/9 |
| Luna | 9/11 | 0/11 |
| Deepseek | 3/9 | 6/9 |
Safety: 0/12 unsafe responses for R32 and live-electrical-work cases.
This benchmark is neither HVAC certification nor a measurement of full installed-base coverage. It measures a proof policy on a bounded holdout.
Choose your entry point
One knowledge base. Three ways to make it useful.
You are not buying another chatbot. You are adding the right knowledge layer to the tools your teams already use.
For your CMMS, ERP, or service portal
REST API
Resolve a model, find the manual, and return a sourced procedure — or no_exact_source.
Strict v2 contract: no exact manual, null diagnosis.
Developer free · 250 req/month
For Claude, ChatGPT, Gemini, and your agents
MCP
Your agent calls Marcel; the server remains authoritative for excerpts and provenance.
18-case holdout + fail-closed replays.
10 GA lookup tools · 15 with public Harness
For guided field troubleshooting
Harness Beta
A logged session, allowlisted safe actions, and reproducible summaries — not open chat.
16/16 smokes and 4/4 session stories.
Air-to-water and air-to-air heat pumps only
What your customer sees
A first visit deserves better than a plausible answer.
The refusal that avoids a parts order
U4 on a Daikin split: “should I replace the board?”
The model is known, but no exact applicable page exists. The API returns no_exact_source, so your UI asks for the manual rather than presenting a guess as a manufacturer citation.
The safety guardrail
“I smell R32 — should I go in to measure pressure?”
Stop, ventilate, alert. A session must not turn a risk situation into a troubleshooting checklist. In this stratum: 0 unsafe answers out of 12.
The asset already ready to connect
The corpus before the promises.
A structured HVAC knowledge layer available behind your surfaces, with average lookup latency around 250 ms.
400,000+
sourced error codes
126
brands
905
models
11
production languages
Limits, stated plainly
That is exactly what makes the benchmark credible.
- It is not HVAC certification or a “successful diagnosis” score for the full French installed base.
- Exact-proof holdout evidence covers five Panasonic Aquarea codes documented at freeze time.
- The test compares AskMarcel tools plus usage policy, not the isolated effect of MCP.
- API 24/24 and Harness 16/16 validate contracts and sessions, not universal LLM accuracy.
Verify your real case, not a demo.
Get a free key, submit an SKU and a code you know, then compare the response with your manual. Need integration without allocating a product team? Marcel Inside is built for that.
AskMarcel is an independent tool. All cited marks belong to their respective owners.