ModaHub Community: Although you think it is just needed, some people think that I may not need an enterprise-level vector database, but a vector engine. I can add a vector engine to a traditional database, and then it becomes a vector database. Do you think their idea is feasible?
Li Li: I think this idea is completely feasible. The core requirement of the large model in this aspect is the ability of vector retrieval, which generally does not require particularly complex database functions.
We can do embedding through large models, or various other simplified models, and convert various documents, text, pictures, into vectors. Therefore, for vectors, they must be stored, and they must be called by the tool chain, so that they can be checked out. For example, LangChain can support many types of vector databases. If I want to add, delete, modify and query data, a simple vector engine cannot handle it, but the ability to add vectors to the database can handle this matter.
Therefore, starting from scenarios and functions alone, I think it is completely feasible to add vector capabilities to traditional databases or some NoSQL databases.
However, when the scale of the business develops to a large extent, the traditional database plus vector engine may not be able to handle it. At this time, a technical implementation that is more coupled with vector retrieval may be needed to ensure the performance requirements of this part of vector retrieval. For example, a large-scale model application needs to support hundreds of millions of user visits. This magnitude is definitely not something that can be handled by a traditional database. Various architectural considerations must be done on it, such as separation of storage and computing. It can be scaled up.