AP-459 Preventing references from crowding out answer in UI #76
Conversation
| <AccordionTrigger>Metadata</AccordionTrigger> | ||
| <AccordionContent> | ||
| <p className="mb-2" style={{ whiteSpace: 'pre-line' }}> | ||
| {props.tind_message || 'no references supplied'} |
There was a problem hiding this comment.
is there an option to not return a References section at all unless props.tind_message is present?
There was a problem hiding this comment.
how are links rendered in this?
There was a problem hiding this comment.
@anarchivist The way that the custom element is called in chatbot/app.py, it should only render the element if a message from langchain has a metadata['tind_message']. Or, after the proposed changes upstream, the element in the data layer should be able to restore it from props.
Once that's changed upstream we may need to alter our chatbot or graph_manager code to make sure the tind_messages are still getting filtered out of the context sent/returned to/from the summarizer.
All that being said, I think the custom_element part of this approach allows us to get a little more specific about how we customize the output of the references. With structured responses or even some advanced string parsing we could add UI signposting like reference counts or links to specific pages in the source documents.
|
It's in Jira, but here's the links to the upstream custom_element issue and PR. |
2 participants
Added a custom_element jsx to put the tind-refs into an accordion component from chainlit's import of shadcn.
There are issues with persisting the element to the data layer because we're not currently using a cloud storage integration.
Cloud storage is required in chainlit's
chainlit_data_layer.pysrc, but it appears to be overly aggressive. If we're not persisting any binary uploads as part of the element, then it should be able to work.Current thoughts are to submit an issue upstream along with a proposed PR for the fix, hookup an S3/localstack integration (that won't be used!), or do some message manipulation on_chat_resume and just recreate them live...