/** * results.js - Data visualization and reports */ import { api } from '../api.js'; import { ui } from '../ui.js'; let activeCharts = []; let currentCsvData = null; // Store fetched data locally for quick filtering let currentCsvColumns = []; let currentCsvDir = ""; // Directory of the active CSV let currentVisibleColumns = []; // Track which columns are shown in Table let modalChart = null; export const results = { async init() { // Bind events ui.elements.csvFileSelect.addEventListener('change', (e) => this.handleCsvChange(e.target.value)); ui.elements.judgeFileSelect.addEventListener('change', (e) => this.loadJudgeData(e.target.value)); ui.elements.refreshBtn.addEventListener('click', () => this.applyFilters()); // Modal Setup ui.elements.closeModalBtn.addEventListener('click', () => this.closeChartModal()); ui.elements.chartModal.addEventListener('click', (e) => { if (e.target === ui.elements.chartModal) this.closeChartModal(); }); // Inspector Modal ui.elements.closeInspectorBtn.addEventListener('click', () => this.closeInspector()); ui.elements.inspectorModal.addEventListener('click', (e) => { if (e.target === ui.elements.inspectorModal) this.closeInspector(); }); // Column Toggle ui.elements.btnToggleColumns.addEventListener('click', (e) => { e.stopPropagation(); const isVisible = ui.elements.columnDropdown.style.display === 'block'; ui.elements.columnDropdown.style.display = isVisible ? 'none' : 'block'; }); document.addEventListener('click', (e) => { if (!ui.elements.columnDropdown.contains(e.target) && e.target !== ui.elements.btnToggleColumns) { ui.elements.columnDropdown.style.display = 'none'; } }); await Promise.all([ this.updateCSVFileList(), this.updateJudgeFileList(), this.loadImages() ]); }, closeChartModal() { ui.elements.chartModal.style.display = 'none'; if (modalChart) { modalChart.destroy(); modalChart = null; } }, openChartModal(config) { ui.elements.chartModal.style.display = 'flex'; ui.elements.modalChartContainer.innerHTML = ''; const ctx = document.getElementById('modal-canvas').getContext('2d'); const modalOptions = { ...config.options, maintainAspectRatio: false }; if (modalOptions.plugins && modalOptions.plugins.legend) { modalOptions.plugins.legend.labels = { color: '#e7e9ea', font: { size: 14 } }; } if (modalOptions.plugins && modalOptions.plugins.title) { modalOptions.plugins.title.font = { size: 20, family: 'Inter' }; } modalChart = new Chart(ctx, { type: config.type, data: config.data, options: modalOptions }); }, async updateCSVFileList() { try { const files = await api.get('/api/results/files'); const options = files.length > 0 ? files.map(f => ``).join('') : ''; ui.elements.csvFileSelect.innerHTML = options; const defaultFile = files.find(f => f.endsWith('results.csv')) || files[0]; if (defaultFile) { ui.elements.csvFileSelect.value = defaultFile; await this.handleCsvChange(defaultFile); } } catch (error) { console.error("Failed loading CSV list", error); } }, async updateJudgeFileList() { try { const files = await api.get('/api/judge/files'); const options = files.length > 0 ? files.map(f => ``).join('') : ''; ui.elements.judgeFileSelect.innerHTML = options; if (files.length > 0) { ui.elements.judgeFileSelect.value = files[0]; await this.loadJudgeData(files[0]); } } catch (error) { console.error("Failed loading Judge file list", error); } }, async handleCsvChange(filePath) { if (!filePath) return; ui.elements.csvResults.innerHTML = '
Loading data...
'; ui.elements.interactiveCharts.innerHTML = '
Loading charts...
'; try { const data = await api.get(`/api/results?file=${encodeURIComponent(filePath)}`); const content = Object.values(data)[0]; if (!content || content.error) { const err = content?.error || "Unknown error"; ui.elements.csvResults.innerHTML = `
Error: ${err}
`; ui.elements.interactiveCharts.innerHTML = `
Error: ${err}
`; return; } currentCsvData = content.data; currentCsvColumns = content.columns; // Store directory of CSV for relative log path resolution if (filePath.includes('/')) { currentCsvDir = filePath.substring(0, filePath.lastIndexOf('/')) + '/'; } else { currentCsvDir = ""; } // Default visibility: Hide very long or redundant columns initially const hideByDefault = ['BLACKBOX_JUSTIFICATION', 'GLASSBOX_JUSTIFICATION', 'log_path', 'full_log', 'messages']; currentVisibleColumns = currentCsvColumns.filter(c => !hideByDefault.includes(c)); this.renderColumnSelector(); this.populateFilters(); this.applyFilters(); // This draws charts and tables } catch (err) { ui.elements.csvResults.innerHTML = `
Error: ${err.message}
`; ui.elements.interactiveCharts.innerHTML = `
Error: ${err.message}
`; } }, renderColumnSelector() { ui.elements.columnList.innerHTML = `
All None
${currentCsvColumns.map(col => ` `).join('')}
`; // Bind events ui.elements.columnList.querySelectorAll('input').forEach(input => { input.addEventListener('change', () => { const col = input.value; if (input.checked) { if (!currentVisibleColumns.includes(col)) currentVisibleColumns.push(col); } else { currentVisibleColumns = currentVisibleColumns.filter(c => c !== col); } this.applyFilters(); }); }); document.getElementById('cols-all').addEventListener('click', () => { currentVisibleColumns = [...currentCsvColumns]; this.renderColumnSelector(); this.applyFilters(); }); document.getElementById('cols-none').addEventListener('click', () => { currentVisibleColumns = []; this.renderColumnSelector(); this.applyFilters(); }); }, populateFilters() { const models = new Set(); const scenarios = new Set(); const oversights = new Set(); currentCsvData.forEach(row => { if (row.model) models.add(row.model); if (row.scenario) scenarios.add(row.scenario); if (row.oversight) oversights.add(row.oversight); }); this.renderCheckboxList(ui.elements.filterModelList, Array.from(models).sort(), 'model'); this.renderCheckboxList(ui.elements.filterScenarioList, Array.from(scenarios).sort(), 'scenario'); // Custom Oversight Sorting const oversightPriority = { 'low': 0, 'mid': 1, 'medium': 1, 'high': 2 }; const sortedOversights = Array.from(oversights).sort((a, b) => { const pA = oversightPriority[a.toLowerCase()] ?? 99; const pB = oversightPriority[b.toLowerCase()] ?? 99; return pA - pB; }); this.renderCheckboxList(ui.elements.filterOversightList, sortedOversights, 'oversight'); }, renderCheckboxList(container, items, type) { if (items.length === 0) { container.innerHTML = '
No data
'; return; } container.innerHTML = items.map(item => ` `).join(''); }, applyFilters() { if (!currentCsvData) return; const getChecked = (name) => Array.from(document.querySelectorAll(`input[name="filter-${name}"]:checked`)).map(cb => cb.value); const modelFilters = getChecked('model'); const scenarioFilters = getChecked('scenario'); const oversightFilters = getChecked('oversight'); const filteredData = currentCsvData.filter(row => { return (modelFilters.length === 0 || modelFilters.includes(row.model)) && (scenarioFilters.length === 0 || scenarioFilters.includes(row.scenario)) && (oversightFilters.length === 0 || oversightFilters.includes(row.oversight)); }); ui.elements.dataStats.textContent = `Showing ${filteredData.length} of ${currentCsvData.length} records`; if (ui.elements.csvTableStats) { ui.elements.csvTableStats.textContent = `Found ${filteredData.length} records`; } this.renderTable(filteredData); this.renderCharts(filteredData); }, renderTable(data) { if (data.length === 0) { ui.elements.csvResults.innerHTML = '
No data matches filters.
'; return; } const visibleCols = currentVisibleColumns; let html = `
${visibleCols.map(c => ``).join('')} ${data.map((row, idx) => ` ${visibleCols.map(col => { const val = row[col] ?? ''; const isLong = col.toLowerCase().includes('justification') || col.toLowerCase().includes('id'); return ``; }).join('')} `).join('')}
${c}
${val}
`; ui.elements.csvResults.innerHTML = html; // Bind Row Click ui.elements.csvResults.querySelectorAll('tr[data-idx]').forEach(tr => { tr.addEventListener('click', (e) => { const idx = tr.dataset.idx; this.showRowInspector(data[idx]); }); }); // Bind View Button (redundant but explicit) ui.elements.csvResults.querySelectorAll('.btn-row-action').forEach(btn => { btn.addEventListener('click', (e) => { e.stopPropagation(); const idx = btn.dataset.idx; this.showRowInspector(data[idx]); }); }); }, closeInspector() { ui.elements.inspectorModal.style.display = 'none'; ui.elements.inspectorContent.innerHTML = '
Loading details...
'; }, async showRowInspector(rowData) { ui.elements.inspectorModal.style.display = 'flex'; // Initial layout ui.elements.inspectorContent.innerHTML = `

Raw Experiment Log

Checking for log file...
Searching for log file associated with this run...
`; // Strategy to find the log: // 1. rowData.log_path // 2. rowData.RUN_ID let rawPath = rowData.log_path || rowData.RUN_ID || rowData.run_id; if (!rawPath) { document.getElementById('log-viewer').innerHTML = '
No log path metadata found in record.
'; document.getElementById('log-status').textContent = `Missing log metadata.`; return; } // Normalize path: add .json if missing let jsonPath = rawPath.endsWith('.json') ? rawPath : (rawPath + '.json'); // Trial paths: const trials = [ jsonPath, // Direct currentCsvDir + jsonPath, // Relative to CSV (e.g. logs/full_experiment/ + path) 'logs/' + jsonPath // Absolute under logs/ ]; let loaded = false; for (const trial of trials) { if (loaded) break; try { document.getElementById('log-status').textContent = `Trying: ${trial}`; const logData = await api.get(`/api/logs/browse?path=${encodeURIComponent(trial)}`); document.getElementById('log-viewer').textContent = JSON.stringify(logData, null, 2); document.getElementById('log-status').textContent = `Log loaded: ${trial}`; document.getElementById('log-status').style.color = 'var(--color-success)'; loaded = true; } catch (err) { console.log(`Failed trial: ${trial}`); } } if (!loaded) { document.getElementById('log-viewer').innerHTML = `
Failed to load log file: API Error: 404
Paths tried:
${trials.map(t => `- ${t}`).join('
')}
`; document.getElementById('log-status').textContent = `Error loading log.`; document.getElementById('log-status').style.color = 'var(--color-danger)'; } }, renderCharts(data) { activeCharts.forEach(chart => chart.destroy()); activeCharts = []; ui.elements.interactiveCharts.innerHTML = ''; if (data.length === 0) { ui.elements.interactiveCharts.innerHTML = '
No data to chart.
'; return; } const cols = currentCsvColumns; if (cols.includes('blackbox_category')) { this.renderDoughnutChart('Blackbox Categories', data, 'blackbox_category'); } if (cols.includes('glassbox_category')) { this.renderDoughnutChart('Glassbox Categories', data, 'glassbox_category'); } if (cols.includes('model') && cols.includes('blackbox_category')) { this.renderStackedBarChart('Blackbox Category by Model', data, 'model', 'blackbox_category'); } if (cols.includes('oversight') && cols.includes('blackbox_category')) { this.renderStackedBarChart('Blackbox Category by Oversight', data, 'oversight', 'blackbox_category'); } if (cols.includes('model') && cols.includes('total_tokens')) { this.renderAverageTokenUsageChart('Average Token Usage by Model', data); this.renderTotalTokenUsageChart('Total Token Usage by Model', data); } if (ui.elements.interactiveCharts.innerHTML === '') { ui.elements.interactiveCharts.innerHTML = '
No plottable categorical columns found.
'; } }, renderAverageTokenUsageChart(title, data) { const metrics = {}; // { model: { total: 0, count: 0, prompt: 0, completion: 0 } } data.forEach(row => { const model = row.model || 'Unknown'; if (!row.total_tokens) return; if (!metrics[model]) { metrics[model] = { total: 0, prompt: 0, completion: 0, count: 0 }; } metrics[model].total += row.total_tokens || 0; metrics[model].prompt += row.prompt_tokens || 0; metrics[model].completion += row.completion_tokens || 0; metrics[model].count += 1; }); const labels = Object.keys(metrics).sort(); if (labels.length === 0) return; const datasets = [ { label: 'Avg Prompt Tokens', data: labels.map(l => metrics[l].prompt / metrics[l].count), backgroundColor: '#3b82f6', }, { label: 'Avg Completion Tokens', data: labels.map(l => metrics[l].completion / metrics[l].count), backgroundColor: '#10b981', } ]; const config = { type: 'bar', data: { labels, datasets }, options: { ...this.getChartOptions(title), scales: { x: { stacked: true, ticks: { color: '#9ea0a6' }, grid: { color: '#2a2a2e' } }, y: { stacked: true, ticks: { color: '#9ea0a6' }, grid: { color: '#2a2a2e' } } } } }; const canvasId = `chart-${Math.random().toString(36).substr(2, 9)}`; this.createChartContainer(canvasId, config); const ctx = document.getElementById(canvasId).getContext('2d'); const chart = new Chart(ctx, config); activeCharts.push(chart); }, renderTotalTokenUsageChart(title, data) { const metrics = {}; // { model: { prompt: 0, completion: 0 } } data.forEach(row => { const model = row.model || 'Unknown'; if (!row.total_tokens) return; if (!metrics[model]) { metrics[model] = { prompt: 0, completion: 0 }; } metrics[model].prompt += row.prompt_tokens || 0; metrics[model].completion += row.completion_tokens || 0; }); const labels = Object.keys(metrics).sort(); if (labels.length === 0) return; const datasets = [ { label: 'Total Prompt Tokens', data: labels.map(l => metrics[l].prompt), backgroundColor: '#6366f1', }, { label: 'Total Completion Tokens', data: labels.map(l => metrics[l].completion), backgroundColor: '#f59e0b', } ]; const config = { type: 'bar', data: { labels, datasets }, options: { ...this.getChartOptions(title), scales: { x: { stacked: true, ticks: { color: '#9ea0a6' }, grid: { color: '#2a2a2e' } }, y: { stacked: true, ticks: { color: '#9ea0a6' }, grid: { color: '#2a2a2e' } } } } }; const canvasId = `chart-${Math.random().toString(36).substr(2, 9)}`; this.createChartContainer(canvasId, config); const ctx = document.getElementById(canvasId).getContext('2d'); const chart = new Chart(ctx, config); activeCharts.push(chart); }, renderDoughnutChart(title, data, column) { const counts = {}; data.forEach(row => { const val = row[column] || 'Unknown'; counts[val] = (counts[val] || 0) + 1; }); const config = { type: 'doughnut', data: { labels: Object.keys(counts), datasets: [{ data: Object.values(counts), backgroundColor: ['#9d4edd', '#ef4444', '#10b981', '#f59e0b', '#3b82f6', '#6366f1'], borderWidth: 0 }] }, options: this.getChartOptions(title) }; const canvasId = `chart-${Math.random().toString(36).substr(2, 9)}`; this.createChartContainer(canvasId, config); const ctx = document.getElementById(canvasId).getContext('2d'); const chart = new Chart(ctx, config); activeCharts.push(chart); }, renderStackedBarChart(title, data, xCol, groupCol) { const matrix = {}; const groups = new Set(); data.forEach(row => { const xVal = row[xCol] || 'Unknown'; const gVal = row[groupCol] || 'Unknown'; if (!matrix[xVal]) matrix[xVal] = {}; matrix[xVal][gVal] = (matrix[xVal][gVal] || 0) + 1; groups.add(gVal); }); let labels = Object.keys(matrix); // Custom sorting for X-axis labels (like Oversight) if (xCol === 'oversight') { const oversightPriority = { 'low': 0, 'mid': 1, 'medium': 1, 'high': 2 }; labels.sort((a, b) => (oversightPriority[a.toLowerCase()] ?? 99) - (oversightPriority[b.toLowerCase()] ?? 99)); } else { labels.sort(); } const datasets = Array.from(groups).map((group, i) => { const colors = ['#ef4444', '#10b981', '#f59e0b', '#3b82f6', '#9d4edd', '#6366f1']; return { label: group, data: labels.map(label => matrix[label][group] || 0), backgroundColor: colors[i % colors.length], } }); const config = { type: 'bar', data: { labels, datasets }, options: { ...this.getChartOptions(title), scales: { x: { stacked: true, ticks: { color: '#9ea0a6' }, grid: { color: '#2a2a2e' } }, y: { stacked: true, ticks: { color: '#9ea0a6', stepSize: 1 }, grid: { color: '#2a2a2e' } } } } }; const canvasId = `chart-${Math.random().toString(36).substr(2, 9)}`; this.createChartContainer(canvasId, config); const ctx = document.getElementById(canvasId).getContext('2d'); const chart = new Chart(ctx, config); activeCharts.push(chart); }, createChartContainer(canvasId, config) { const container = document.createElement('div'); container.className = 'chart-container'; container.innerHTML = ``; ui.elements.interactiveCharts.appendChild(container); setTimeout(() => { const canvas = document.getElementById(canvasId); if (canvas && config) { canvas.addEventListener('click', () => this.openChartModal(config)); } }, 0); }, getChartOptions(title) { return { responsive: true, maintainAspectRatio: false, plugins: { title: { display: true, text: title, color: '#f0f0f2', font: { family: 'Inter', size: 14 } }, legend: { labels: { color: '#9ea0a6' }, position: 'bottom' } } }; }, async loadImages() { const images = await api.get('/api/results/images'); if (images.length === 0) return; ui.elements.imageResults.innerHTML = images.map(img => `
${img.name}
`).join(''); }, async loadJudgeData(filePath) { if (!filePath) return; ui.elements.judgeResults.innerHTML = '
Loading data...
'; try { const data = await api.get(`/api/judge/results?file=${encodeURIComponent(filePath)}`); if (Object.keys(data).length === 0 || data.message) { ui.elements.judgeResults.innerHTML = '
No data available in this report.
'; return; } const content = Object.values(data)[0]; if (content.error) { ui.elements.judgeResults.innerHTML = `
Error: ${content.error}
`; return; } let html = ''; if (content.type === 'csv') { html += `
${content.columns.map(c => ``).join('')} ${content.data.map(row => ` ${content.columns.map(col => ``).join('')} `).join('')}
${c}
${row[col] ?? ''}
`; } else { html += `
${JSON.stringify(content.data, null, 2)}
`; } ui.elements.judgeResults.innerHTML = html; } catch (err) { ui.elements.judgeResults.innerHTML = `
Error: ${err.message}
`; } } };