AI assistant for adjusting Dandiset metadata

[magland]

From suggestion of @satra, exploring the possibility of a chat interface for setting Dandiset metadata. Could take care of some of the more painstaking aspects such as extracting authors from a publication, target brain structures derived from uberon, or extracting information from the dandiset.

I made a test page to show how this could be done
https://neurosift.app/?p=/test&dandisetId=000722

with a sample Dandiset 000722,

You can talk to that assistant on the main page, but it can't yet edit things directly (but it can probe information about the Dandiset and knows its limitations).

But you can use a chat to edit the Contributors.

Click on the "Contributors" link in the right panel. Then ask it to do things like:

"Remove Chiara" (sorry Chiara if you ever read this :))

"Correct spelling errors" (there actually is a typo in this dataset)

"Add authors from this citation [paste]"

"Please reformat the names to conform to last, first"

[satra]

nice! also weird to see awardnumber associated with a person.

would we do it for just contributors or show a diff of any change in any section? i.e. should we treat the editable metadata as a whole. there are some parts of the metadata that are read only, and the json schema knows this. those parts are only edited by the backend service. users cannot change them.

[magland]

I was imagining that you'd be able to edit each section... but then you'd make all the changes in a single operation once ready (that's where the dandi API request would happen).

[magland]

But idk, maybe it's better to save each section individually. on the fence.

[satra]

i don't think there is a clear notion of which UX is better. just that it would be good to also identify sections that have no information.

[magland]

Another thought. I am putting this as a test page on neurosift just for convenience. But really this should be a standalone app since it does a very specific thing. It could eventually be absorbed by DANDI... but idk, pros/cons.

[satra]

even more ideally it should just be an additional ui element of the current meditor. perhaps people can switch to this "mode" versus the editing mode.

[magland]

Regarding the schema and parts that are readonly, etc. I've been going off of this spec

https://github.com/dandi/dandi-archive/blob/master/web/src/types/schema.ts

Is there a better place I should be looking for this? A json or yaml schema?

[satra]

https://github.com/dandi/schema/tree/master/releases/0.6.8

[satra]

that's the latest release.

[magland]

Okay cool. So for the readonly fields... do they just get ignored when updating metadata via api call?

[satra]

yes. the server discards/overwrites them as appropriate.

[magland]

@satra I am now using the provided dandiset.json and feeding that as context to the assistant (the entire thing is around 10k tokens, but I extract just the part needed for contributors and it's more like 2k tokens).

To test this you can prompt it with something like "Add a fake organization with a lot of sample info to highlight the schema"

[magland]

And it also does validation as part of the tool call to guarantee that it is valid.

[magland]

@satra - now you can (almost) save data back to the dandiset.

Different link:
https://neurosift.app/?p=/import

You can put in your DANDI Staging API key

It will show the staging dandisets owned by you. I have added you as an owner to 215759 to use as an example.

Select a dandiset.

Now you can use the chat interface to modify the contributors.

A button appears "Save changes to Dandiset Metadata"

Clicking that almost works to save it back to dandi staging.

The issue I believe is that CORS access needs to be added for the domain neurosift.app (localhost:3000 would also be nice). The reason I think that CORS is the only issue is that when I "Copy as curl" from the dev console of the browser and paste it into a terminal, it works.