This topic describes the data backflow features, cost information, and other details.
Data backflow supports collecting and storing data generated in the environment, such as input requests and prediction results, for later optimization. Continuous data generated in the environment can be filtered or optimized through data backflow tasks and used for subsequent data analysis or as higher-quality data sources.
For example, in AI Agent business scenarios, it is necessary to monitor users’ actual requests and key data nodes during processing, to uncover hidden issues affecting results. Online data can also be sent back as needed, including end-to-end input and output, LLM input and output, and data used to supplement evaluation or training sets to continuously improve AI Agent performance.
Data backflow tasks created by TLS are essentially data transformation and log shipping tasks. When you create a data backflow task, the system automatically generates an associated data transformation task. Based on the preset time window and selected backflow fields, it processes the specified range of log data and saves the processed results (containing only the selected backflow fields) to the designated log topic. This supports scenarios such as data analysis, log aggregation, and information filtering.
If TOS is selected as the target, the system automatically creates a TOS shipping task in addition to the data transformation task. For details on TOS shipping tasks, see Overview of shipping logs to TOS. Data backflow tasks support the following data sources and target options:
- If the data backflow target is a TLS log topic, costs are the same as for a standard log topic.
- If the data backflow target is set to ship logs to AgentLoop evaluation sets and TOS, costs include TLS log shipping fees and TOS storage fees. For details, see Overview of shipping logs to TOS.