Quickstart

Quickly get familiar with Evolution's capabilities

Choose observability, evaluation, or optimization for your current task. These capabilities can be used independently; you do not need to complete them as three sequential stages.

Observability

Evolution provides developers with full-chain, visualized observability of the execution process, completely recording every processing stage from input to output. It also supports monitoring statistics, rate-limiting statistics, and alert rule configuration.

Minimum preparation: Select a published, supported application in the current workspace. Observability supports Agent 1.0, Agent 2.0, and workflow applications, but not agents created through the Assistant API. For first-time use, open Application Observability Configuration to authorize the OpenTelemetry service role, activate the service, and initialize its LogStore; see Observability prerequisites. Use the Alibaba Cloud primary account; RAM users need the required permissions granted by that account.

Click Enable Observability for the target application, submit a test question to generate trace data, then open its details and Trace List to check inputs, outputs, and nodes. Data synchronization requires observability to be enabled for that application.

1. Report and view call-chain traces and model / tool events.
2. Observe monitoring metrics such as request count, tokens, models, and tool invocations.
3. View rate-limiting metrics such as application QPM and throttling counts.
4. Configure alert rules based on alert templates and review alert history.

For more details, see Observability and Alert Management.

Evaluation

Evolution provides developers with systematic evaluation capabilities to assess application outputs and agent execution trajectories. By analyzing experiment results, you can dig into edge cases and make targeted improvements.

Minimum preparation: Prepare input samples, expected results, a corresponding dataset, and scoring dimensions. Publish the dataset before use; drafts cannot be used for evaluation. In the task, select the published dataset version and target application. Map every grader variable to dataset fields or application outputs with matching data types; see Grader mapping.

1. Create and manage evaluation datasets, organizing input samples and expected results.
2. Create or select a grader, configuring rule- or model-based scoring dimensions.
3. Create evaluation tasks and view results—run in batch to generate scores, and analyze score distribution and regression comparisons.

For more details, see Evaluation Tasks.

Optimization

Evolution provides developers with a systematic evaluate-and-optimize loop. Based on debugging results, evaluation dataset data, and online trace feedback, you can optimize application prompts intelligently or manually, compare versions side by side, and adopt the best result in a single click—continuously improving application output quality.

Minimum preparation: Prepare the application's current prompt and a test question that exposes the issue. Check that the application has been created and published, as recommended in Prerequisites. For debugging and feedback optimization, submit the test question, select the corresponding debugging result, describe the issue and desired improvement, then click Start Optimization. For an evaluation-data path, prepare the chosen dataset version or completed evaluation task results; an online trace path needs existing observation feedback.

1. Compare multiple prompt versions to debug and optimize your application prompt.
2. Optimize your application prompt based on debugging results and human feedback.
3. Optimize your application prompt based on high-quality evaluation datasets and evaluation task results.

For more details, see Optimization.

Next steps