Skip to content

CLAUDE.md

Latest commit

 

History

History
70 lines (50 loc) · 2.57 KB

CLAUDE.md

File metadata and controls

70 lines (50 loc) · 2.57 KB

Frontend CLAUDE.md

This file provides guidance for working with the frontend visualization dashboard.

Frontend Architecture

The frontend is a Dash-based web application for interactive visualization of causal attribution graphs. It displays feature relationships, cluster analysis, and intervention results.

Core Components

Main Application (app.py):

  • Dash app initialization and layout setup
  • Integrates all visualization components

Data Layer (data/):

  • loaders.py: Loads attribution graphs, clustering data, and intervention results
  • models.py: Data models for graph nodes, edges, and metadata

Visualization Layer (visualization/):

  • graph/: Core graph rendering (nodes, edges, layout, renderer)
  • components/: Reusable UI components (cluster manager, feature display, etc.)

Callbacks (callbacks/):

  • graph_callbacks.py: Interactive graph updates and filtering
  • cluster_callbacks.py: Cluster analysis and intervention controls
  • annotation_callbacks.py: Feature annotation and explanation handling

Key Data Flow

  1. Graph Loading: data/loaders.py loads attribution_graph.pt files containing:

    • sparse_pruned_adj: Pruned adjacency matrix from attribution analysis
    • Feature indices, token information, and intervention results

  2. Node Processing: Active nodes determined by adjacency matrix row sums

  3. Edge Processing: Edges filtered by weight thresholds and active nodes

  4. Layout: Networkx-based graph layout with clustering support

  5. Rendering: Plotly-based interactive visualization

Common Issues

Nodes with No Outgoing Edges:

  • Attribution graph correctly prunes these nodes via feature_mask
  • Frontend should use saved feature_mask instead of recalculating from adjacency
  • Check data/loaders.py lines 104-105 for active mask calculation

Graph Filtering:

  • Multiple threshold controls: feature influence, edge weights, clustering
  • Ensure consistent filtering across all visualization components

Configuration

Settings (config/settings.py):

  • Graph layout parameters, color schemes, filtering thresholds
  • Cluster analysis parameters and intervention settings

Development Commands

Launch frontend:

python src/circuit-tracing-visual-interface/frontend/launch.py

Data Requirements

Expected data structure in attribution_graph.pt:

  • sparse_pruned_adj: Pruned adjacency matrix
  • feature_indices: Feature location indices
  • input_tokens, token_string: Input text data
  • top_logit_*: Prediction information
  • Optional: feature_mask for proper node filtering