In real life, if two people have a conversation, it generally requires a three-step process: one party first throws out the topic as an introduction; the other party first mobilizes memory to judge whether he understands the topic, and then analyzes and gives what answer he should make. This cycle goes on and on until the interaction ends, and this dialogue will be absorbed by both parties as a new "memory".
In order to allow the computer to complete such an interactive process and continue to become daily in one-to-one or one-to-many situations, AI scientists have proposed a CVP structure, namely "ChatGPT (large model represented by ChatGPT) + Vector Database (vector database) Database) + Prompt (prompt words)", respectively undertake the functions of computer analysis, memory and introduction.
As a general existence of computer memory, vector database is attracting the attention of a large number of investors and entrepreneurs. Xie Chao, founder and CEO of Zilliz, a vector database start-up, told Jiemian News that when implementing a large model, an important reality to face from the perspective of data is the separation of computing and storage, that is, the large model belongs to the manufacturer, while the data belongs to the user. "Almost all mainstream large-scale model manufacturers in China came to us for cooperation in the first half of the year, and they were eager to know one thing - how to use large-scale models with vector databases, or how to combine computing and storage to achieve low-cost reuse. "
A vector database is a new type of database that specializes in processing (mainly including storage and retrieval) unstructured data. Traditional databases mainly deal with structured data stored in two-dimensional tables of rows and columns. This type of data has a standardized format and is easier to do quantitative analysis. Unstructured data refers to high-dimensional, difficult-to-quantify abstract data that usually requires a specific data structure to organize and is not easy to analyze. In real life, unstructured data appears in various forms, including text, images, audio and video, and in the future, multi-modality will present more complex and diverse expressions, postures, etc.