Alibaba Cloud Summit | Application Practice of Unified Recall Engine in Search Scenarios

AI Online Serving Engineering System


Alibaba Cloud Summit self-developed complete search engineering system, the AI ​​Online Serving system, currently supports all search, recommendation, and advertising businesses of Alibaba's e-commerce at home and abroad. The technology backbone of China and Taiwan, AI·OS is already the infrastructure of the whole group including e-commerce, Alibaba Cloud, Youku, Cainiao, Hema, Dingding, etc. More importantly, the cloud products of the AI·OS system ( Open search and intelligent recommendation) matrix serves global developers through Alibaba Cloud Summit and is industry-leading in terms of stability and engineering efficiency.
Unified Recall Engine
Unified engine architecture and evolution process
The left picture shows the different execution processes of the search engine HA3 and the recommendation engine BE. We abstract the functions of each engine into operators, and form the basic functions into a public operator library. Users can directly reuse and develop according to business needs, forming the Suez on the right. frame.
Alibaba Cloud Summit Features of the Unified Recall Engine
1. DAG the query process
• Unified with deep learning execution engine
• The search function is abstracted into operators
• Unified operator library to support reuse and development of operator granularity
2. Alibaba Cloud Summit Multiple query expressions
• SQL
• TuringSDK
Wait.....
The execution process can be flexibly customized to accelerate the speed of business iteration
Application Practice of Unified Recall Engine
Challenges of recalling engines
Alibaba Cloud Summit want, want, want
1. Data inflation: document data, algorithm data
2. The application of deep learning: recall, rough sorting, fine sorting
3. Stable and efficient: high availability, timeliness, low latency
Traditional solutions and problems
The expansion of data scale is reflected in the increasing number of data dimensions. For example, in the field of e-commerce search, only two dimensions of merchants and commodities were considered before, but now we also need to consider dimensions such as logistics and location. The traditional engine processing joins these data offline into a large wide table and pushes it to the online index construction and query service. This will cause a problem. It is likely that an auxiliary table data update will lead to a large number of main table data updates, resulting in write data. The problem of expansion poses a great challenge to the timeliness of online services, which is difficult to meet in some scenarios, especially in big promotion scenarios, which are difficult to meet the needs of low latency and high timeliness.
Traditional solutions:
Alibaba Cloud Summit The data is split according to a certain dimension to provide services through multiple engine instances, and the business side splits a query into multiple requests to access multiple engines to achieve search results.
Existing problems:
1. Serialization of a large amount of data occurs;
2. Data may be truncated, resulting in impaired performance;
For example, in a takeaway platform search, it is found that the store you want to search is not matched due to delivery time or distance, resulting in the intentional search menu not being reflected, and the user experience being poor;
Another manifestation of data scale expansion is that the amount of data increases, which leads to more time for a single search load to provide a query.
Traditional solutions:
Alibaba Cloud Summit one is to expand the index, which may lead to the splitting of requests and the merging of results. As the number increases, it takes more and more time, and gradually becomes a technical bottleneck. Another is that when the number of searches is large, the stability and availability of the entire cluster are compromised, and query results are unstable for users.
Unified Recall Engine Solution
1. The engine supports multiple tables
• Load multiple tables online at the same time through one engine, and the index construction, update, switching, and loading of each table are independent;
• Through the online multi-table join method when querying, you can get global information in one query, including store information and product information, can be fully utilized to match the recall results that best meet the needs of users;
2. Use SQL to express the query process
• Easy for developers to use
• Reuse basic functions of SQL ecosystem
3. Parallel query, a tool to reduce latency
The index data is divided according to a certain dimension, and the parallel query can be divided according to different divisions when processing the user's query request, thereby reducing the delay of the entire query and avoiding the problems caused by the expansion method.
4. Vector recall, deep learning is applied in the recall stage
In today's information-rich world, it is difficult for our query engine to meet the needs of business by relying on text query alone.
• Adopt the vector retrieval kernel-Proxima developed by DAMO Academy, with the construction of super-large-scale data vector index, providing high-performance online vector retrieval capability;
• On the basis of the original text recall, adding vector recall can achieve both the document recall rate and the accuracy rate, and at the same time, better and flexible configuration can be performed in each sorting, and good search results can be achieved
Application of unified recall engine in recommendation scenarios
Create a recall engine for personalized recommendation effects
Cloud practice of unified recall engine
Alibaba Cloud Open Search
Alibaba Cloud Summit openSearch is a one-stop intelligent search business development platform based on Alibaba's self-developed large-scale distributed search engine. It provides a fully open engine with built-in capabilities such as query semantic understanding and machine learning sorting algorithms in various industries. ability to help developers quickly build intelligent search services with higher performance and higher search baseline effects.
Application of open search in e-commerce industry
• E-commerce industry search productization has been implemented, users do not need to explore technology in all directions, but only need to access according to the template to have better search services;
• Built-in higher-quality algorithm model, eliminating a lot of data labeling and model training work, and directly built-in Tao-based search algorithm capabilities;
• Support personalized search and service capabilities, and realize important services such as search results, drop-down prompts, and shading words through the multi-channel recall capability on the engine side;
• Support user-trained NLP models to be imported into Open Search to flexibly meet the needs of business developers;
• Alibaba's self-developed engine system can handle massive data, high concurrency, and massive user requests, with better performance than open source solutions;
• According to the changes of e-commerce banks, iteratively update the original capabilities and provide more timely service guarantees;
Application of Open Search in Educational Search Scenarios
• Support text index, image vector index, and formula index to recall results in multiple ways, reducing the ineffectiveness of text search questions and photo search questions;
• A full set of capabilities for educational query analysis to solve the problem of low accuracy, customizable sorting scripts, and in-depth optimization of the sorting effect of recall results;
• Alibaba Cloud SummitVector + text recall flexibly configured by users to quickly improve the effect of the search system;
• Sorting plug-in development - Cava language, stronger customization ability, easier maintenance, easy to achieve business sorting requirements;
•Alibaba Cloud Summit Pay-as-you-go, with immediate effect, ensuring stable search during peak periods, without the need to purchase a large amount of resources in advance, without cost burden;
• It supports millisecond-level response for search of hundreds of billions of data, and real-time data updates are visible in seconds;

Related Articles

Explore More Special Offers

  1. Short Message Service(SMS) & Mail Service

    50,000 email package starts as low as USD 1.99, 120 short messages start at only USD 1.00