As the creator of Milvus, the world's most popular open source vector database, Zilliz provides next-generation database technology for AI applications to help organizations easily develop AI applications.
With the mission of democratizing AI, Zilliz is committed to simplifying AI data management infrastructure and empowering more enterprises with vector databases.
Co-founder Guo Rentong said in an interview: the intersection of vector databases and large models is mainly knowledge augmentation, and the mainstream programs in the past year are still loosely coupled, i.e., the results of the knowledge base recalled are filled into the prompt to do the augmentation of the model inputs. In the future, there will be some tightly coupled forms.
One is that the big model will be deeply involved in the selection of knowledge, especially after the cost of long context big model continues to decline, the input of background knowledge can be greatly widened, and through the big model in the depth of the semantic level of knowledge selection. From a technical point of view, this knowledge selection process has two layers, the bottom layer is supported by traditional indexing, and the top layer is supported by Attention. The new technology will make the area covered by Attention move down, which also means the effect will be improved.
Another direction is the semantic space fusion of embedding model and LLM. This direction is not so clear at the moment, but it is quite imaginative. After the fusion of these two semantic spaces, the knowledge fragments in the vector database can directly participate in the Attention process. This is a more natural form of "big model memory", which can support dynamic knowledge selection. For example, using the data representation inside the model as the input to the memory, an extended Attention action is performed in the memory to recall the content that is strongly related to the current context. Of course, this direction still has great challenges in terms of model structure and cost.