Extracting information from a large database can be difficult, especially when you don't know what to look for. That's exactly why ChatGPT and friends were born to assist Google. However, these models are not to be used for personal data. Sometimes, you just want to know what you did last week, or last year. Sometimes, you just want to know what kind of person you were. And, sometimes you want a deep analysis on your thought and past. Here is how Memoiz works. First, you have your primary database. This is where you store all your text data. Then, you might have a secondary database. This is where you store the embeddings of your text data. (It is possible that both can be on Redis Stack) When the text data is "archived", its content will be embedded by Cohere Embed, turning into a (seemingly but not) meaningless 4096-dimension vector saved into the Redis server. When the user starts chatting with Memoiz, the chat history will determine the keywords for semantic search on the Redis server. The closest vectors will be treated as a context for the text generation prompt. The text generation prompt will then be sent to the Cohere Generate model for the chat response. In the next steps, we plan to develop a larger-scale mood tracker with a more powerful language model. We also plan to share a public API for developers to use Memoiz in their own projects.Category tags:
"Super original idea. Good use of tech and connections to DB set up. I would love to see you build it further "