How Mia works
Mia is the assistant inside Dlubal CALC. Between your message and the answer sit three things worth knowing about, because each one explains a behaviour that otherwise looks arbitrary.
The response language
Mia first considers your explicit language request or the language of your message. Short confirmations and form submissions retain the conversation language. If it is unclear, an explicit language preference in your enabled memories can help. The app language is the final fallback. The response language can differ from the interface language.
The tier decides how hard it thinks
Two tiers, and you choose in the message bar:
| Tier | For |
|---|---|
| Standard | Most questions. One pass, fast |
| High | Long chains, several interacting checks, anything with an attachment |
A Standard message costs 1 credit, a High one 2. Every new chat starts on Standard; a High choice applies to the conversation you made it in and is not carried into the next one.
Some turns are pinned to High whatever you picked. An attachment is the clearest case: the Standard model may not be able to read the file at all, so answering with it would produce a confident answer about a document nobody looked at.
Substantive answers in voice conversations
Voice Agent handles the spoken conversation and passes substantive questions to CALC. Even with OpenAI GPT-Live, those answers use the selected Standard/High model. Choosing a voice provider in Settings does not replace your model choice for calculations and engineering questions.
Tools run when the answer needs them
Mia does not answer everything from memory. When a question needs a rule, a location or a model, it calls a tool and the result goes into the answer:
- the Dlubal knowledge base and your company knowledge, for technical questions and internal documents
- geo data for a location, covering snow, wind, seismic, climate and altitude
- web search, when you allow it
- RFEM, when it is connected
Dlubal-operated cloud tools remain available and do not need global switches. Availability is not execution. Mia calls a tool only when the question needs it. RFEM remains switchable because it is a local application where Mia can read or change a model.
The connection status in Settings checks whether CALC can reach the available tools. A successful check does not yet confirm a result for a particular location. Automatically retrying a failed status check does not repeat a tool action.
You can also see the time of the fresh connection check and the last recorded Geo call with its outcome and duration. Call history covers up to 24 hours for your current access. Missing history does not establish a connection failure. The display does not start an additional Geo query.
For RFEM, the capabilities offered by the Connector also determine what Mia can do. A connection can work even when writing or Python is not permitted. Mia should state that limit rather than claim a connection failure. See Working with RFEM.
Costs apply only to the actions listed in Plans and credits, and only when they actually run. That is why two similar questions can cost different amounts.
Example: "what is qL²/8" costs one message and nothing else. "Which snow load applies in Passau" costs the message plus one geo lookup, because it has to fetch data for the location.
It says what it assumed, and asks when it cannot
The rule Mia follows is to state assumptions rather than bury them, and to ask rather than invent when something essential is missing. That is why an incomplete prompt gets a question back instead of a plausible number.
There is a second rule worth knowing because it shapes answers about the software itself: Mia will not tell you that a setting is on or off by default, or where exactly a control sits, unless a retrieved source said so. If you need a default and no source states it, you get told to check it in the program. A wrong default sends you hunting for a checkbox that was already ticked.
Result diagrams and verification summary
When a calculation supplies structured result data, CALC displays the corresponding cards below the answer. If requested results are missing, CALC requests a correction within the available calculation budget. Incomplete data produces a notice; missing curves are not estimated from the answer text.
For skills with a fixed calculation component (for example the compression member comparison), CALC calculates after the input form is submitted and shows the cards from that result. Engineering rules such as the wind assumption or the stiffness reserve then live in the skill and in the result, not in general instructions to the model.
Equilibrium of frames and beams is checked by a calculation, not taken from the answer text. If an answer states that equilibrium is satisfied but no such check in the same answer passed, CALC asks Mia once to run the check and correct the values. If that is not possible, a note below the answer says that equilibrium is not verified or that the check failed; treat the reactions and internal forces as unverified in that case.
The selected deflection unit and precision apply to new answers and structured deflection diagrams. Unit changes convert displayed values while preserving calculation data. See Settings.
Sources below the answer
The source list is not filled with additional search hits when Mia cited no source. Language variants of the same Dlubal article are combined. When a translation in the conversation language is verified, the link opens that version; otherwise, the available language version is retained.
What it can see
This conversation, the files attached to it, and the results of the tools it called during it. Not your other chats, not a colleague's, and nothing on your machine that you did not attach or expose through the Dlubal-Connector.
Where to go next
- Citations for what a source reference means.
- Web and geo data for the two lookups you can steer.
- Limitations for where to be careful.