Search Answer Streaming for Long-Tail Query Response Time
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing information search systems face challenges in providing timely and accurate results, especially for long tail queries, due to the slow processing times of machine learning models and reduced recall rates, leading to unsatisfactory user experiences.
Innovation Solution
A method where a server determines if a search request is associated with a predetermined category, generates and sends a partial answer result using a machine learning model, and subsequently obtains and sends the remaining content through streaming protocols to enhance user experience and improve recall rates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If the machine learning model processes the entire search request to generate complete answer results, then the accuracy and completeness of search results are improved, but the search time consumption increases significantly
Solution Approach 1:
The patent segments the answer result into multiple parts (first part and second part) and transmits them separately. The first part containing key information is sent immediately, while the second part is sent subsequently, allowing users to receive useful information quickly without waiting for complete processing
Solution Approach 2:
The system performs preliminary classification to determine whether the search request belongs to a predetermined category before invoking the machine learning model. This preliminary action allows the system to optimize processing paths and reduce unnecessary computational overhead for non-predetermined category requests
2Loss of information
If the system waits for the machine learning model to generate complete answer results before sending them to the client, then the completeness of information is improved, but the timeliness of information delivery deteriorates
Solution Approach 1:
The answer result is divided into multiple segments transmitted at different times. The first part is sent immediately after generation, providing timely information, while the second part follows later to complete the information set, balancing timeliness and completeness
Solution Approach 2:
The system maintains continuous information delivery by sending the first part of the answer result without interruption, then continuing to send the second part. This continuous action ensures users receive information progressively rather than experiencing a complete waiting period
3Reliability
If the system processes all search requests through the machine learning model, then the recall rate is improved, but the processing efficiency and productivity deteriorate
Solution Approach 1:
The processing system is segmented into different paths: predetermined category requests undergo machine learning model processing to maintain high recall rate, while non-predetermined category requests use alternative processing methods, improving overall processing efficiency
Solution Approach 2:
The system applies the machine learning model only partially - specifically to requests matching predetermined categories - rather than processing all requests uniformly. This selective application maintains high recall rate for relevant queries while improving overall system productivity
Data Source
AI summary
Embodiments of the present disclosure relate to a method, an apparatus, a device, a medium and a program product for searching information. The method comprises in response to determining that request content in a search request from a client is associated with a predetermined category, obtaining a result identifier of an answer result for the search request, and generating, by a machine learning model, content of the answer result for the request content. The method further comprises sending a first response comprising the part of content to the client to display the part of content of the answer result, and in response to receiving from a client an obtaining request, obtaining the remaining part of content generated by the machine learning model, and sending a second response comprising the remaining part of content to the client for displaying the answer result.