Topic

Build a customized in house VectorDB at Pinterest

Since 2023, Pinterest has been internally designing and implementing tools and systems to serve large language models, including VectorTableSearch, Text-To-SQL in QueryBook, and automated table documentation.Pinterest uses the industry's Pinterest uses industry generic GPT models, such as openAI. To increase the flexibility of the system and to add Pinterest's own features, we make extensive use of Retrieval Augmented Generation (RAG) to add Pinterest-specific elements to the large model, such as engineering documentation, query history, and automated table documentation. We heavily use Retrieval Augmented Generation (RAG) to add Pinterest-specific elements to the larger model, such as engineering documentation, query history, and Slack chats. In this process, we faced many challenges, such as the lack of standards for training data, which makes it difficult to reuse. Lack of VectorDB development and operation and maintenance, and each team often works separately. Inability to perform continuous optimization of large models, and so on. To solve the appeal problem, we designed and developed VectorDB as a service within Pinterest, including a unified data import platform for the company, support for data indexing and updating, and high-performance query services. This talk will detail the VectorDB service within Pinterest, and the difficulties in customer service during the process of building VectorDB. Translated with DeepL.com (free version)