I’ll try to find them… it would be a good set of references at least.
Yes. Perhaps there could be preferences/permissions for which functions the AI is allowed to use, whether a cloud-based AI should see only masked data for living people, etc.
We had the same discussion for webtrees. See the chapter “Security” at GitHub - Jefferson49/webtrees-API: A webtrees 2.2 custom module to provide a Web based API, also supporting MCP · GitHub
Webtrees as a multi-user system has a role based user model. So it is possible to define granularily the rights of an user that is used by an AI. So we are able to hide information about living persons. The other important concept in webtrees is that it is possible to make changes in the genealogical data that need a review before they are released. So we can allow the AI to change data; but a human reviewer can check them before they are integrated.
It would be great if this assistant would know about the role of relations within a tree and could make consistency checks and would give an answer to e. g. the following question: Are there any conflicting or possible conflicting or missing data in the tree regarding the statement “person X is the son of parents Y and Z”?
Background: In the earliest church records in Germany there are often many people having identical names and only rare additional information like “is the son of” or is the “youngest” (at this concrete time). Without the knowlegde about the time before the church records start it’s a hard job to regenerate all relations between the persones.
How would you do this manually using Gramps? If you can describe what you do, then we could give the assistant tools to do the same.
It’s the job, I currently try to do manualy. Because I just found by fortune (nearly at the end of the analysis of the anaylsing church records for 2 small villages) a contradiction within the growing tree by adding the death of a woman: as wife of a family she died as widow - but the current selected husband died later. The mother was added to this family because he was called in the death record of one son. Because there are several men having 100% at this time the same name as the selcted husband, at least the combination of the persons within this family, esspecially of the father/husband, is wrong by the new information.
But I stop here, I see my mistake by myself: There is no standardized information nor the meaning of an information like “wife died as widow” direct available for any assistent.
I dreamed about a global evidence check at the end using such an assistant. But without enough information within the data this cannot be possible.
I’ve been working to make Gramps Assistant work well with smaller models. It takes some special handling, but it will be worth it. For example, a large number of tools can overwhelm a smaller model. I’ve developed a system that will only bring in the tools dynamically, as needed. Still pure Python; no dependencies.