EdgeChains
EdgeChains copied to clipboard
eval/feedback system
edgechains must have built in support to capture feedback for prompts/final answers (just like for token counts which we already do)
- User feedback: thumbs up/down associated to a piece of data that I can trace back to a workflow run that used a specific version of a prompt/template
- Output metrics: certain thresholds that flag LLM-generated output (in my case it's related to ecommerce search; so if something turns up high on search results but low on click/conversion it could be a telling me that my prompt is generating bad data)