Hi all-
I’ve built a custom GPT called Budget Wine Snob, and I’d really appreciate some honest feedback from Berserkers.
Its purpose is to help identify genuinely good wine bargains from flash-sale sites, retailers, and winery-direct offers. The goal is not just to list cheap bottles, but to surface deals that seem truly worthwhile and to be clear about what looks well-supported versus what may be a bit more uncertain.
If you have a few minutes to check it out, I’d love to hear your impressions. I’d be especially interested in whether it feels useful, credible, and easy to follow, and whether the recommendations and explanations seem sensible.
Please feel free to share any reactions, critiques, or suggestions for improvement. I’d really value your thoughts.
This is a just for fun thing, and I make no money off of it.
I found it thoroughly frustrating and totally unhelpful, but I may not be your target audience. It would require quite a bit of prompting to reliable be a useful tool for me, at which point I would just have my own LLM.
i queried for some deals on low intervention burgundy producers and while the wines recommended fit the bill, the suggestions were a bit obvious for someone with only a modest amount of knowledge. the real problem was hallucinated wine.com links and completely inaccurate price quotes.
Not sure what is going on at that site, as I haven’t done any business over there. For what it’s worth, my GPT is just looking for retail stuff, not auction sales as far as I know, and it does not interact with those sites in terms of placing bids or anything.
i think it will be difficult to completely eliminate numerical discrepancies in output. in order to handle numbers with precision the model would need to use the code execution tool. otherwise it will rely upon the learned weights to output a statistically probable (but incorrect) number. also in order to get real time data the model would have to use the function calling tool to access an external api. perhaps be explicit with your instructions on how to fetch the data, evaluate it, and output into structured format such as json. i just attempted a similar exercise using a normal gpt window and got a much better (but still not perfect) response.
Here is a suggestion. Focus on leveraging AI to achieve the goals of various “bargain hunter/great deals” threads, without requiring users to spend time reading through all of them. Essentially, the AI would monitor recommendations made by knowledgeable wine enthusiasts and apply a personalized filter, trained on your preferences, to highlight only the offers that align with your interests.
From a technical standpoint, this would involve a vector database where documents are metadata-tagged for each user. A straightforward RAG (Retrieval-Augmented Generation) implementation could then retrieve content matching a user’s documented preferences.
I tell my wine-buying AI agent that I like this wine or this type of wine, and here are my general budget constraints (this becomes a metadata-tagged document that is associated with me and trains my AI model). My AI agent later tells me that “John Doe” recommends this wine for this price, and I reply that I think John Doe has terrible tastes and would never buy a wine they liked unless it was a low alcohol wine from Burgundy, where the person actually seems to align with my palate. That also becomes a document, metadata tagged to me, and used by the AI when analyzing offers to my preferences. From now on, the AI would filter recommendations using all those criteria.
Over time, the model would learn increasingly nuanced preferences, such as “That was not a bad suggestion, but I avoid buying from ‘WINE SHOP’ due to a previous negative experience. However, I might consider their offers if they feature exceptional deals on German Riesling.”
The AI could also send entertaining chat messages such as: “Hi Jim, I know you think John Doe has the taste of a yak and buys things you’d never drink, but he found a Burgundy from a producer you like in an excellent vintage, and the price seems well below market, so maybe the guy actually discovered something good. Here’s the link. Let me know what you think.”
