Chats

A chat is one piece of work. Everything in it shares context: the system you described, the assumptions you corrected, the files you attached, the results that came back. That is why the second question understands the earlier exchange while a new chat has to be told everything again.

One chat per task, not per day

Start a new chat when the subject changes. A chat that has been a beam, then a slab, then a question about credits carries all of it into every later answer. That occupies context and gives the model three subjects to confuse; it does not change the fixed chat price.

Example: you finished a purlin check and now want to look at the bracing. New chat. You want the same purlin in S355 instead of S235? Same chat, one sentence.

A new chat always starts on the Standard tier. If the previous one ran on High, that choice stays with it; pick High again in the message bar when the new subject needs it.

How much context goes into the next answer

CALC selects conversation context automatically. There is no fixed user setting such as "10 turns". The latest completed messages go first. If the content becomes too large for the selected model's context window, the oldest messages are left out of that request. The latest exchange is retained.

This trimming applies only to the individual model request. It does not delete messages from the chat. You can still read and search older content or add it to a report.

Example: early in a long chat, you explained a load assumption. Many large tool results later, that explanation no longer fits in the next model request. Repeat the decisive assumption in a new message if it must still govern the next check.

Stopping an RFEM request

Keep the browser open when stopping an RFEM request. After the stop is confirmed, CALC saves your request, the partial answer, and the tool calls observed by the browser. If saving fails, CALC shows an error. Once saved, the request remains marked as stopped after reloading. This does not protect against a closed tab or browser crash.

Calls that already started may have changed the model. If a result is missing, the outcome remains unknown. Have the current model state inspected before requesting further changes. CALC does not automatically replay the stopped request. After a failed RFEM write, further writes in the same execution loop are blocked; read-only inspection remains available.

Improving dictated text

Voice Input automatically detects the spoken language by default and keeps the transcript in that language, independently of the CALC interface language. If recognition needs help, select a specific dictation language in Settings or return to Detect automatically.

After dictating, choose Improve text from the input field's three-dot menu. The provider and model are shown there before you click. The edit uses the configured Standard-tier model to improve grammar and punctuation in the original language. It starts only when you click and can be undone from the same menu while the edited draft remains unchanged. This feature is unavailable in local dictation mode.

Finding one again

The sidebar lists your chats newest first, grouped by date, and the search above the list matches titles and message text. Chats are titled automatically from the first exchange, so the list stays readable without you naming anything.

A separate short background call creates the title. CALC uses a fast, low-cost model for this. This choice does not change the model that answers your actual question. If that model is overloaded or does not answer within a few seconds, CALC tries the fast model of another provider once. Only if that fails too is a shortened first line of your message used as the title. The title is written in the language of your message; if a short message does not show it clearly, in your CALC interface language.

How long they are kept

Retention is part of the plan, and it is the one limit that removes work rather than blocking it:

What a chat can see

The conversation, the files you attached to it, and whatever the tools looked up during it. Not your other chats, not a colleague's chats, and nothing on your machine that you did not attach or expose through the Dlubal-Connector.

This is worth knowing when an answer seems to have forgotten something: if it happened in a different chat, it did not happen as far as this one is concerned.

Reading an answer aloud

The speaker icon below an answer reads it aloud. Units, standards, and short formulas are spoken as words, so "2.5 kN/m²" becomes "2.5 kilonewtons per square meter."

On plans from Starter up, Mia's own voice reads the answer. Each month includes free minutes; after that, each answer read aloud costs 1 credit, and playing the same answer again is free. Very long answers are read up to about six minutes, and the rest stays in the chat. On the Free plan, or whenever Mia's voice is unavailable, your browser's voice reads the answer instead.

Starting Voice Agent

When starting Voice Agent, a new chat is saved first. If saving fails, the voice session does not start. You can try starting again; a failed save remains retryable. If the chat is still unavailable, reload CALC and open an available chat.

For chats with voice usage, the latest answer shows the credits consumed by the whole session, including chat and the voice conversation. The tooltip beside the number shows both parts. Earlier answers retain their own chat charge; the session total appears only once. Later settlements are refreshed periodically. Missing billing records are not presented as free usage.

Saved result cards

Structured result diagrams and verification summaries are saved with the answer. They remain available when you reopen the chat and in its shared read-only view. Older answers without this result data do not gain retrospectively calculated diagrams.

Use the PDF button in a result diagram dialog to add individual diagrams or all diagrams in the scrolling view to the report. Diagrams from the same answer share one chapter and fill each PDF page before continuing onto the next. Previously created separate chapters remain unchanged.

RFEM model views

In the RFEM viewer, you can save up to 20 named views with the camera, visible objects, selection and section box in the chat. Restore them when reopening the chat or in its shared read-only view. If saving fails, the viewer indicates that changes have not been synchronised and offers a retry.

Reading the same RFEM model again can highlight added, changed and removed objects relative to the previous import. After a successful chat save, the last differing import remains available as a comparison baseline after reload and in shared read-only views. This is not a full revision history.

Sharing

A chat can be shared as a read-only link from its menu in the sidebar. The reader sees the conversation and its results without an account of their own.

Where to go next

• Projects to group chats and fill the report cover page.
• Reports to turn a chat into a document.
• Plans and credits for the retention and limit table.