Principle:Marker Inc Korea AutoRAG Results Dashboard
| Knowledge Sources | |
|---|---|
| Domains | Visualization, RAG Pipeline Optimization |
| Last Updated | 2026-02-12 00:00 GMT |
Overview
The results dashboard provides interactive visualization of optimization trial results, enabling human review and comparison of module performance across pipeline nodes.
Description
After an optimization trial completes, the raw results exist as CSV summaries and parquet DataFrames scattered across the project directory structure. While these files are machine-readable, they are difficult for humans to interpret without tooling. The results dashboard bridges this gap by presenting trial outcomes through an interactive web interface.
The dashboard organizes information into multiple tabs. The Summary tab displays a markdown-rendered overview of every node's best module, including its name, configuration parameters, and metric values. Each node tab provides detailed views including a summary DataFrame with all candidate modules and their scores, strip plots and box plots showing metric distributions across candidates, and a drill-down view that lets users inspect individual module result DataFrames row by row. The Used YAML file tab shows the exact configuration that produced the trial results, ensuring full reproducibility.
This design supports the iterative nature of RAG pipeline optimization, where practitioners run multiple trials with different configurations and need to compare outcomes efficiently. The tabular and visual presentation makes it easy to spot patterns such as modules that perform consistently well across metrics versus modules that excel on one metric but fail on others.
Usage
The results dashboard is used after an optimization trial has completed. It is typically launched from the command line or programmatically, pointed at a specific trial directory. It serves as the primary tool for human analysis of trial results and for deciding whether to adjust the configuration for a subsequent trial.
Theoretical Basis
The dashboard follows a read-only, hierarchical presentation model:
Level 1 -- Trial Summary: Aggregates the best module selection from every node across all node lines into a single markdown document. For each node, it shows:
- The best module name
- The module's parameter configuration
- A table of metric names and their values
Level 2 -- Node Detail: For each node, displays:
- A tabular summary of all candidate modules with their metric scores, execution times, and a boolean flag indicating the selected best
- Statistical visualizations (strip plots and box plots) of metric distributions across candidates
- An interactive drill-down that loads the full result DataFrame for any selected module
Level 3 -- Configuration Reference: The raw YAML configuration rendered as a code block, providing a direct link between the results and the experiment definition.
The visualization stack uses the following rendering pipeline:
trial_dir
-> summary.csv (trial-level)
-> node_line_name/node_type/summary.csv (node-level)
-> node_line_name/node_type/*.parquet (module-level)
Panel Tabs:
Tab "Summary" -> Markdown pane from make_trial_summary_md()
Tab "{node_type}" -> Column layout with Tabulator + Matplotlib plots
Tab "Used YAML" -> Markdown pane from yaml_to_markdown()
The dashboard uses the Panel library for layout and interactivity, Bokeh for table formatting, and Matplotlib with Seaborn for statistical plots. It runs as a local web server on a configurable port, making it accessible through any web browser.