LangSmith experiment page is slow to load or not render fully with large datasets

The LangSmith experiment results page experiences slow loading times and may crash the browser when dealing with large datasets, particularly those containing substantial outputs like images. This performance issue can be mitigated by utilizing the display settings to hide the main output column and selectively display only the necessary output fields as individual columns.

Issue Description

Users have reported performance degradation when viewing the experiment results page in LangSmith for experiments with large datasets. The primary symptoms include:

• Extremely slow loading times for the experiment results.

• The browser becoming unresponsive or crashing completely.

• These issues are particularly prevalent when the outputs of the experiment contain large data, such as images.

Environment

• Product: LangSmith

• Version: any

• Area: Experiment Results View

Cause

The root cause of the slow performance and browser crashes is the overwhelming amount of data being loaded and rendered in the experiment results view. When an experiment's output column contains large objects, such as high-resolution images, the browser struggles to handle the data transfer and rendering, leading to the observed issues. In one instance, image data amounted to ~500 MB, which is substantial for a web interface to manage.

Workaround

To prevent the slow loading and crashing, you can customize the Display settings of the experiment view to avoid rendering the large output data directly.

Resolution

The issue can be resolved by adjusting the display settings in the experiment view to manage how data is presented. This involves hiding the comprehensive "output" column and instead, creating separate columns for specific fields of interest from the outputs.

Step by Step Instructions

1. Navigate to the experiment results page in LangSmith.

2. Locate the "Display" settings, typically found in the header of the experiment view.

3. Within the display settings, you can configure which columns are visible.

4. Hide the main "output" column. This will prevent the browser from attempting to load the large image data within that column.

5. Add new columns by extracting specific, necessary fields from the outputs. For example, if your output contains an image and some metadata, you can create a new column to display only the metadata.

6. This customized view will load significantly faster as it is only handling smaller, more manageable pieces of data, and the rows will load incrementally as you scroll.

References

https://docs.langchain.com/langsmith/analyze-an-experiment