I’m developing a Custom GPT designed to act as a virtual bartender, pulling cocktail recipes from a wide range of sources, including renowned cocktail books and social media influencers. However, I’m encountering unexpected hallucinations in the GPT’s output, despite it being fed accurate data via Actions by calling AirTable.
tl;dr: A GPT-based bartender that curates recipes dynamically from AirTable but is showing data inconsistencies.
Issue: The GPT seems to be jumbling its data, leading to incorrect recipe suggestions. Has anyone experienced similar issues? What could be causing this, and how might it be resolved? (See photos)
Alt: First, GPT Interface without Action does not know what the cocktail “Green Eyes” is.
Second, AirTable screenshot showing the ingredients and instructions on how to make the cocktail “Green Eyes” that is Gin based and garnished with a lime wheel and cherry.
Third, GPT Builder interface with Action hooked up and already digested all of the AirTable records where it knows the cocktail “Green Eyes” with almost accurate specs, but thinks it’s a Rum drink and not a Gin drink and thinks the garnish is zest and not a cherry. Some of the other measurements are a little off as well.