问HN：RAG可以用于推荐系统吗？
我已经为这个项目工作了一周，但很难明确需求。我们的想法是创建一个完全自定义的内容推送（具有隐私保护！没有任何数据会发送到服务器），跨不同平台。以下是我遇到的一些问题：
1. 我正在使用sentence-transformers/all-MiniLM-L6-v2来提高速度（人们绝不会愿意为单个页面等待几秒钟），而且每个“聚类”都是根据历史的点赞、观看和不喜欢情况以及我自定义的布局进行评分。我不确定的是，用户是否会接受我的布局（目前我将HN转变为类似TikTok的滚动体验）。
2. 在没有任何数据的情况下，很难为评分系统设定基准。我对当前基于算法的平台感到不满，因为它们往往忽视个人口味。它们推荐的内容是基于与我口味“相似”的“其他”用户的偏好。此外，我们无法选择“探索”级别。我认为在X平台上，只有10%到20%的内容对用户来说是“新”的，而这个比例用户自己完全无法调整。因此，我无法从这些网站获取更多探索性的内容。
这两个问题让人感到困惑，但我真的不知道该如何继续，因为目前没有现成的应用或工具可以解决这个问题。如果你有任何需求，请分享给我 :)
我非常需要你的反馈，以判断我是否应该继续这个项目！
查看原文
I have already worked on this for a week, but it's very hard to define the needs.
The idea was to create a fully customized feed (with privacy! nothing is sent to the server), across different platform. Here are some questions that I have encountered:<p>1. I'm using sentence-transformers/all-MiniLM-L6-v2 for speed (people will never wait seconds for a single page), and every "clusters" are scored based on historical likes/watch/dislike with my custom layout. The thing that I'm not sure here is whether people will accept my layout(for now I was transforming the HN into tiktok-like scrolling experience).
2. Without any data, it's very hard to calibrate the scoring system for the baseline. One thing that I am uncomfortable with current algo-based platforms is that they are rather ignorance on personnal taste. They recommend based on what "others" similars to my taste instead. Anothing is that we can not select the "exploring" level. I think only 10~20% of the feeds on X are "new" to the user, and this proportion is not adjustable at all by the users themselves. However, I am thus not able to get more exploring contents from these websites.<p>It's quite massy with these two questions, but I really don't know how to continue since there are no existing apps/tools for this at all. If you have any needs, please share it :)<p>Really need your interaction to see if I should continue on this project or not!!!