AI analysis
Identify interfaces, categorise requirements by an outline profile, suggest features – in a working copy, with evidence and only after your approval.
Work in progress. This documentation is not complete yet. If something is missing, write to us: support@wiestigo.com
As of 11 October 2026 · checked against Wiestigo Mirador 0.1.0
The AI analysis makes suggestions about your requirements. Every suggestion names what it is based on, and none of it changes your project before you have applied it and written it back.
The AI analysis is a licence feature, each analysis separately. An analysis your licence does not include is listed but cannot be chosen.
The analyses
| Analysis | what it suggests |
|---|---|
| Identify interfaces | one folder per interface with the requirements concerning it, and an overview of the interfaces |
| Categorize requirements by outline profile | exactly one chapter of an outline profile for every requirement (default: DI-IPSC-81431A), and the chapter folders for it |
| Suggest features from the text | features from your feature trees for every requirement. It only adds and never removes a feature; it also needs variant management. |
Start an analysis
Every analysis works in a working copy, never directly in the project.
- In the AI analysis tab, choose the Analysis.
- Create a New working copy – with a name and an Entry point in the tree – or open an existing one. Optionally Mirador locks the scope in the project while you work.
- Choose the Scope (the whole working copy or selected folders) and the options of the analysis.
- Start analysis. Progress is shown; Cancel run stops it.
Anyone who may edit in the project may start one.
Review the proposal
Under Review the proposal, every suggestion is shown with its evidence:
- the text passage it is based on,
- the source – learned knowledge or the language model – and a confidence,
- uncertain, if the analysis will not commit itself.
A language-model interface assignment whose passage cannot be found in the text is discarded – with a notice. Such a chapter assignment stays, marked uncertain. Check feature suggestions yourself against the passage they quote.
You deselect, rename, merge interfaces or set a different chapter. Apply writes your selection into the working copy; Discard drops it.
Back into the project
Only Write back to RMM changes your project. If someone else has changed an affected requirement in the meantime, that place is skipped and named – unless you deliberately choose to overwrite other people’s changes. Afterwards you can discard the working copy.
What stays local and what leaves your network
For categorising by outline profile and identifying interfaces, Mirador first uses learned knowledge that runs locally in the service. It learns from assignments you have written back and stores weights, not requirement texts. Weights learned from only a few examples can, however, allow conclusions about those requirements – so treat exported knowledge as confidentially as the texts. Only where this knowledge is not certain enough does Mirador ask a language model – if one is configured. Suggest features, by contrast, sends every requirement in scope to the language model if one is configured.
In your own data centre, whether and which language model takes part is an operations setting: none (local analysis only), a service via the Anthropic API, or a service with an OpenAI-compatible interface, including a model you run yourself. Only this is sent to the language model:
| Analysis | sent |
|---|---|
| Identify interfaces | id, title and text of the requirements, the names of known interfaces |
| Categorize | id and text of the requirements, the chapters of the profile |
| Suggest features | id, title and text of the requirements; qualified name (TREE.FEATURE), label, description and kind of the features |
If the language model fails, Mirador carries on locally and says so in the result. A daily token budget per user can be set; by default it is unlimited. How to set this up in your own data centre is described under Self-hosting.
We operate the Wiestigo cloud in the EU, except for language-model services, some of which we source from US providers and none of which you have to use.