Create New Dataset

Steps to Create a Dataset

1. Navigate to the Dataset Page

   • After creating a new project, you will be automatically redirected to the dataset page.

2. Start with an Empty Screen

   • If no datasets exist, you will see a guided instruction screen prompting you to create your first dataset.

3. Create a Dataset Using Tracing

   • Datasets in agentic applications are created by committing a trace.

   • Refer to the Tracing Documentation to learn how to set up and capture traces.

4. Commit Your Trace

   • Once a trace is successfully recorded and committed, it will automatically populate the dataset list on the UI.

   • Each dataset represents a specific experiment conducted within the application.

5. View the Dataset

   • After committing a trace, the dataset will appear in the UI, allowing you to access and manage it as an experiment.

---

Dataset Columns:

1. TraceID

2. Timestamp

3. Trace URL

4. Feedback

5. Response

6. Metadata

7. Metric Score

8. Completion Token

9. Prompt Token

10. Cost

---

Trace View and Prediction Details

When you click on a datapoint in the dataset table, the Trace View opens, providing detailed insights at the span level:

Trace View

• Displays all the information instrumented during tracing, broken down at the span level.

• Offers a granular view of the execution flow, including:

  • LLM calls.

  • Tool interactions.

  • Agent decision-making steps.

• Enables debugging and performance analysis by identifying bottlenecks or errors.

Prediction Details

• Users can toggle to the Prediction Details tab to focus on the evaluation metrics.

• Key features include:

  • Metric Evaluation Score: A breakdown of scores based on pre-defined metrics such as Response Correctness, Context Relevancy, etc.

  • Reasoning: Detailed explanations behind each score, offering insights into why a particular rating was assigned.

---