Docs

Torch Log Service

AgentLoop

Quick Start with AgentLoop

In this topic, the integration of the VeADK large model observation application is used as an example to show how to quickly use AgentLoop for log services, from data collection to completing experiment monitoring.

Prerequisites

The source project and log topic have been created, with key-value indexing configured.

The target project and log topic have been created, with indexing configured.

Integrate VeADK log application

Select the region where log service is located from the top navigation bar.

In the left navigation bar, select AgentLoop > LLM applied observability.

On the LLM app tab, click Integrate application on the VeADK card.

Configure the integration application parameters, and then click Integrate application.

Parameters	Description
Name	The name of the log application to be integrated.
Description	Description information for the log application.
Region	The region of the log application to be integrated.
Log project	Select the project where data will be stored.
Report center	Select the report to be generated, such as veadk Observability in this example, and then set the following parameters: Dashboard name: Set the dashboard name corresponding to the report. Log topic: Choose a log topic as the data source. Field validation: The system automatically matches the relevant index fields in the log topic based on the log application type. You need to replace the fields manually if the followings happen: some fields are missing, the field types differ, or statistics are not enabled. Dashboard field description. For details, see Log field description for preset dashboard.

After successful integration, you will be directed to the dashboard page by default.

Ship VeADK application data to review set

Select the region where log service is located from the top navigation bar.

In the left navigation bar, select AgentLoop > LogBack Flow.

On the LogBack Flow page, click Create LogBack task.

On the LogBack task creation page, follow the prompt instructions to complete the configuration.

Enter the basic configuration.

Configuration	Description
Task name	Task name for the data backflow task. This name will automatically be used as the name for the scheduled SQL task.
Description	A brief description of the data backflow task.

Configure the data source.

Configuration	Description
Data source type	Select the data source type for the data backflow task: currently, only Torch Log Service TLS is supported.
Log project	The project containing the original logs for the data backflow task.
Log topic	The topic containing the original logs for the data backflow task.

Configure filter rules.

Configuration	Description
Backflow field	Backflow fields are sourced from key-value fields in the data source index. Custom fields are supported to add. You can use the original field name or set an alias (aliases can not be duplicated) as the exported field name. You must set field aliases in the following scenarios: JSON format fields Reserved fields in TLS
Condition configuration	You can filter results by setting conditions for callback fields and support multiple AND relationship filters.

Write data processing rules.

The system automatically generates AgentLoop transformation rules. It is not recommended to edit these rules directly.

If you need to add other transformation rules, follow the onscreen instructions to enter data transformation rules in the input box. TLS supports a variety of DSL syntax functions. On the transformation functions tab, you can click a function name to view syntax descriptions and examples, and insert specified syntax with one click to quickly write transformation rules. For the list of supported syntax functions, see Functions overview.

Test data processing rules.

Follow the page instructions to test by using log data from the log topic or specified test data. Click Execution preview to test processing rules.

Configure scheduling rules.

Configure scheduling rules for LogBack Flow tasks.

Configuration	Description
Start time	The scheduling start time of the log backflow task, i.e., the time when the first instance is scheduled to start.
End time	The scheduling end time of the log backflow task.

Set the target.

Configuration	Description
Target Region	The target region where the log of data backflow is located.
Log project	The target project where the log of data backflow is stored.
Log topic	The target topic where the log of data backflow is stored. The target topic can not be the same as the source log topic.
Other targets	Select AgentLoop review set and configure the following parameters: Evaluation method: Select Create new evaluation dataset. Target review set name: Enter the name of the review set. Field mapping to review set: Select the corresponding review columns for the original fields.

Click Submit.

Create experiment to complete evaluation

Select the region where log service is located from the top navigation bar.

On the left navigation bar, select AgentLoop > review > experiment.

On the Experiment page, under the Experiment list tab, click Create experiment.

Complete the following configurations on the experiment creation page according to the interface prompts.

Select the review set.

Parameters	Description
Review set	Select the review set.
Version	Select the review set version.
Maximum concurrent executions	Set the maximum number of concurrent executions. Valid values: 1–20. The experiment supports concurrent execution of items in the review set, but it is limited by the concurrency of the evaluation object and the model TPM of the evaluator. This is the ideal maximum number of executions.

(Optional) Select evaluation target

Currently, configuring the evaluation target is not supported. You can skip this step.

(Optional) Add evaluator

If you skip evaluator configuration, the experiment will only obtain the Agent's actual output. This is suitable for scenarios with manual labeling.

Click Add evaluator and configure the following parameters:

Parameters	Description
Name	Select an evaluator.
Version	Select the version of the evaluator.
Prompt details	View prompt details.
Field mapping	Mapping of review set fields and evaluated object outputs to evaluator fields. This allows accurate input collection for evaluation.
Maximum concurrent executions	Set the maximum number of concurrent executions. Valid values: 1–20. The experiment supports concurrent execution of items in the review set, but it is limited by the concurrency of the evaluation object and the model TPM of the evaluator. This is the ideal maximum number of executions.

Confirm experiment configuration

Enter the experiment name and confirm the experiment configuration. After verification, click Start experiment.

View evaluation data

After starting the experiment, the system will automatically perform the evaluation task. Once the task is complete, you can manually refresh the page to view detailed data.

Last updated: 2026.07.30 18:21:21