This paper introduces you to the application of vector retrieval services in scenarios such as e-commerce intelligent search and preference recommendation, AI Q&A systems such as natural language processing, multimodal search in gallery-type websites, video retrieval, molecular detection and screening, and so on.
E-commerce intelligent search and preference recommendation scenarios
In the e-commerce intelligent search and preference recommendation scenarios, vector databases can realize the search and recommendation functions based on vector similarity. For example, an e-commerce platform contains images and description information of various commodities, and when users search for commodities, they can query related commodities through the image or description information, and they also want to realize the recommendation function, which automatically recommends the commodities that may be interested in to the users.
The user only needs to first convert the images and description information of the commodities into vector representations using Embedding technique and store them in the vector database. When the user inputs a query request, the vector retrieval service may convert it to a vector representation, then calculate the similarity between the query vector and all the merchandise vectors in the vector database, and then return several merchandise vectors with the highest similarity. In addition, the historical browsing records and purchase records of a user may be converted into vector representations based on the user's historical behavior and preferences through the vector retrieval service, and the commodity vectors that are most similar to the vectors as well as those with higher similarities may be queried in the vector database, so as to recommend commodities that may be of interest to the user, and to provide smarter and personalized services, and more efficient and excellent performance and purchasing experiences.
Natural language processing and other AI Q&A system scenarios
Q&A systems are common real-world applications belonging to the field of natural language processing. Typical Q&A systems such as Tongyi Thousand Questions, ChatGPT, online customer service system, QA chatbot, etc. For example, in a Q&A system, which contains some predefined questions and corresponding answers. The user wants to be able to automatically match the most similar predefined questions based on the input questions and return the corresponding answers. In order to realize this feature, firstly the predefined questions and answers can be converted into vector representations by a vector retrieval service and stored in a vector database. Secondly when the user inputs a question, the vector retrieval service can convert it to a vector representation and query the vector database for the question vector that is most similar to that vector. Then use the steps of model training, Q&A reasoning, and post-optimization to achieve a language intelligent interaction system similar to Tongyi Thousand Questions, ChatGPT, and so on.
Multi-modal search scenarios for gallery-type websites
Current large-scale image material sites and sharing social applications, etc., usually have hundreds of millions or even tens of billions of images, and can only provide simple text search or a single way to search for images, so that users can not quickly find the images they need. Instead, using the DashVector vector search service, the image content and text descriptions are represented as vectors and stored in a vector database. When the user searches, it supports various modal search modes such as text search, graph search and combined text + image search with precise filtering, etc. The search requirements are also represented by vectors, and similar searches are carried out in the vector database, which helps the user to find the desired image quickly, thus improving the user experience.
Video Retrieval Scenarios
In video retrieval scenarios, such as video surveillance systems, movie and television resource websites, short video applications, and other platforms, which contain a variety of video data. Using vector retrieval service by converting the video data into vector representation and storing it in a vector database. When a user sees a movie clip or a video screenshot, the video similarity search system is used to perform content vector-based video retrieval so that the most similar video to the query video can be quickly retrieved and returned to the user as the search result. A clustering-based video retrieval method can also be used in the vector database to cluster videos and perform fast retrieval between clusters to improve retrieval efficiency and accuracy.
Molecular detection and screening scenarios
In molecular detection, molecular fingerprints (e.g., ECFP, MACCS bonds, etc.) can be used to convert molecular structures into vector representations and store them in a vector database. When the user inputs a query request, the same method can be used to convert it to a vector representation and query the molecular vectors that are most similar to that vector in the vector database and return it to the user as the search result, realizing molecule retrieval and screening based on molecular structure similarity. It provides more intelligent and efficient solutions for molecular discovery and drug design.