Jump to content

Connect SuperML | Leeroopedia MCP: Equip your AI agents with best practices, code verification, and debugging knowledge. Powered by Leeroo — building Organizational Superintelligence. Contact us at founders@leeroo.com.

Principle:Marker Inc Korea AutoRAG Trial Summary And Dashboard

From Leeroopedia
Revision as of 17:46, 16 February 2026 by Admin (talk | contribs) (Auto-imported from principles/Marker_Inc_Korea_AutoRAG_Trial_Summary_And_Dashboard.md)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)


Knowledge Sources
Domains Visualization, Evaluation
Last Updated 2026-02-08 06:00 GMT

Overview

A results presentation technique that aggregates evaluation metrics and provides an interactive dashboard for analyzing RAG pipeline trial outcomes.

Description

After an evaluation trial completes, AutoRAG generates CSV summary files containing the best module per node along with their metrics and execution times. The dashboard module provides an interactive web interface (built with Panel/Bokeh) for exploring trial results. It displays a trial summary tab, per-node detail tabs with strip/box plots of metric distributions, and a tab showing the original YAML configuration. This enables visual comparison of module performance across different metrics.

Usage

Use load_summary_file to programmatically access trial results. Use dashboard.run to launch an interactive web dashboard for visual analysis of a completed trial.

Theoretical Basis

The summary aggregation follows a hierarchical rollup pattern:

  1. Each node produces a summary.csv with per-module metrics and execution times
  2. The node summary identifies the best module (is_best flag) based on the strategy
  3. The trial summary aggregates best modules from all nodes into a single overview
  4. The dashboard renders this hierarchy as navigable tabs with visualizations

Related Pages

Implemented By

Page Connections

Double-click a node to navigate. Hold to expand connections.
Principle
Implementation
Heuristic
Environment