This question requires an in-depth discussion of the relationship between large models and vector databases. This issue has aroused our thinking since the launch of ChatGPT last year. At the time, we were keenly aware that this would be an opportunity. However, in China, it took longer for this concept to be recognized. I personally noticed during my trip to the United States in April and May last year that databases have become a very hot topic in the United States, but they do not seem to have been widely recognized in China. Therefore, we need to start with the relationship between vector databases and large models.
In my opinion, this relationship solves two important problems in the field of artificial intelligence: generation and retrieval. In the beginning, many people debated whether domain knowledge could be gained by fine-tuning large models. Is it appropriate to combine vector databases with large models? Now more and more opinions show that fine-tuning mainly brings some regular cognition, and real knowledge needs to be supplemented by external storage. This gave rise to the need for a vector database.
In the above context, we proposed a concept called "CVP Stack", where "V" stands for vector database, "C" stands for ChatGPT, and "P" stands for prompt. In our view, large models represent reasoning capabilities, vector databases complement its knowledge, and prompts represent control logic to implement specific semantics. The combination of these three can achieve excellent business results.
Since vector databases play a storage role between large models and storage, multiple requirements for vector databases arise. First, large models need to store massive amounts of data