mstoeck3/forensic-pathways
2
0
Fork 0
Code Issues 3 Pull Requests Projects Releases Activity

forensic-ai #4

Merged
mstoeck3 merged 20 commits from forensic-ai into main 2025-08-05 20:56:02 +00:00
Conversation 0 Commits 20 Files Changed 18 +22 -79
6 changed files with 22 additions and 79 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files
Showing only changes of commit b515a45e1e - Show all commits

7
src/components/AIQueryInterface.astro
View File

@ -702,12 +702,10 @@ class AIQueryInterface {
toolsByPhase[phase] = [];
});
// DEBUG: Log recommendation structure
console.log('[AI Results] Recommendation structure:', recommendation);
console.log('[AI Results] Recommended tools:', recommendation.recommended_tools);
recommendation.recommended_tools?.forEach(recTool => {
// DEBUG: Log each tool's confidence data
console.log('[AI Results] Tool confidence data:', recTool.name, recTool.confidence);
if (toolsByPhase[recTool.phase]) {
@ -716,7 +714,7 @@ class AIQueryInterface {
toolsByPhase[recTool.phase].push({
...fullTool,
recommendation: recTool,
confidence: recTool.confidence, // Ensure confidence is passed
confidence: recTool.confidence,
justification: recTool.justification,
priority: recTool.priority,
recommendationStrength: recTool.recommendationStrength
@ -836,13 +834,11 @@ class AIQueryInterface {
return '';
}
// Calculate summary statistics
const totalTime = auditTrail.reduce((sum, entry) => sum + entry.processingTimeMs, 0);
const avgConfidence = auditTrail.reduce((sum, entry) => sum + entry.confidence, 0) / auditTrail.length;
const lowConfidenceSteps = auditTrail.filter(entry => entry.confidence < 60).length;
const highConfidenceSteps = auditTrail.filter(entry => entry.confidence >= 80).length;
// Group entries by phase for better organization
const groupedEntries = auditTrail.reduce((groups, entry) => {
if (!groups[entry.phase]) groups[entry.phase] = [];
groups[entry.phase].push(entry);
@ -1048,7 +1044,6 @@ class AIQueryInterface {
second: '2-digit'
});
// Reuse existing grid and text utilities
return `
<div class="border-l-2 pl-3 py-2 mb-2" style="border-left-color: ${confidenceColor};">
<div class="flex justify-between items-center mb-1">

8
src/config/prompts.ts
View File

@ -2,7 +2,6 @@
export const AI_PROMPTS = {
// Main tool selection prompt
toolSelection: (mode: string, userQuery: string, selectionMethod: string, maxSelectedItems: number) => {
const modeInstruction = mode === 'workflow'
? 'Der Benutzer möchte einen UMFASSENDEN WORKFLOW mit mehreren Tools/Methoden über verschiedene Phasen. Wählen Sie 15-25 Tools aus, die den vollständigen Untersuchungslebenszyklus abdecken.'
@ -51,7 +50,6 @@ Antworten Sie NUR mit diesem JSON-Format:
}`;
},
// Scenario analysis prompt
scenarioAnalysis: (isWorkflow: boolean, userQuery: string) => {
const analysisType = isWorkflow ? 'forensische Szenario' : 'technische Problem';
const considerations = isWorkflow ?
@ -74,7 +72,6 @@ ${considerations}
WICHTIG: Antworten Sie NUR in fließendem deutschen Text ohne Listen, Aufzählungen oder Markdown-Formatierung. Maximum 150 Wörter.`;
},
// Investigation approach prompt
investigationApproach: (isWorkflow: boolean, userQuery: string) => {
const approachType = isWorkflow ? 'Untersuchungsansatz' : 'Lösungsansatz';
const considerations = isWorkflow ?
@ -96,7 +93,6 @@ ${considerations}
WICHTIG: Antworten Sie NUR in fließendem deutschen Text ohne Listen oder Markdown. Maximum 150 Wörter.`;
},
// Critical considerations prompt
criticalConsiderations: (isWorkflow: boolean, userQuery: string) => {
const considerationType = isWorkflow ? 'kritische forensische Überlegungen' : 'wichtige methodische Voraussetzungen';
const aspects = isWorkflow ?
@ -179,7 +175,6 @@ WICHTIG:
- "pros" soll die Stärken für diese spezifische Aufgabe hervorheben`;
},
// Background knowledge selection prompt
backgroundKnowledgeSelection: (userQuery: string, mode: string, selectedToolNames: string[], availableConcepts: any[]) => {
return `Wählen Sie relevante forensische Konzepte für das Verständnis der empfohlenen Methodik.
@ -200,7 +195,6 @@ Antworten Sie AUSSCHLIESSLICH mit diesem JSON-Format:
]`;
},
// Final recommendations prompt
finalRecommendations: (isWorkflow: boolean, userQuery: string, selectedToolNames: string[]) => {
const prompt = isWorkflow ?
`Erstellen Sie eine Workflow-Empfehlung basierend auf DFIR-Prinzipien.
@ -225,7 +219,6 @@ WICHTIG: Antworten Sie NUR in fließendem deutschen Text ohne Listen oder Markdo
}
} as const;
// Type-safe prompt function with proper overloads
export function getPrompt(key: 'toolSelection', mode: string, userQuery: string, selectionMethod: string, maxSelectedItems: number): string;
export function getPrompt(key: 'scenarioAnalysis', isWorkflow: boolean, userQuery: string): string;
export function getPrompt(key: 'investigationApproach', isWorkflow: boolean, userQuery: string): string;
@ -238,7 +231,6 @@ export function getPrompt(promptKey: keyof typeof AI_PROMPTS, ...args: any[]): s
try {
const promptFunction = AI_PROMPTS[promptKey];
if (typeof promptFunction === 'function') {
// Use type assertion since we've validated the function exists
return (promptFunction as (...args: any[]) => string)(...args);
} else {
console.error(`[PROMPTS] Invalid prompt key: ${promptKey}`);

7
src/pages/index.astro
View File

@ -265,20 +265,15 @@ const phases = data.phases;
return;
}
// AI Button click handler
if (aiQueryBtn) {
aiQueryBtn.addEventListener('click', () => {
// Visual feedback
aiQueryBtn.classList.add('activated');
setTimeout(() => aiQueryBtn.classList.remove('activated'), 400);
// Switch to AI view
switchToView('ai');
// Trigger existing view change system
window.dispatchEvent(new CustomEvent('viewChanged', { detail: 'ai' }));
// Scroll to AI interface
if (window.scrollToElementById) {
window.scrollToElementById('ai-interface');
} else {
@ -294,14 +289,12 @@ const phases = data.phases;
const filtersSection = document.getElementById('filters-section');
const noResults = document.getElementById('no-results');
// Hide all views first
if (toolsGrid) toolsGrid.style.display = 'none';
if (matrixContainer) matrixContainer.style.display = 'none';
if (aiInterface) aiInterface.style.display = 'none';
if (filtersSection) filtersSection.style.display = 'none';
if (noResults) noResults.style.display = 'none';
// Show selected view
switch (view) {
case 'grid':
if (toolsGrid) toolsGrid.style.display = 'block';

65
src/utils/aiPipeline.ts
View File

@ -63,17 +63,15 @@ interface AnalysisContext {
auditTrail: AuditEntry[];
// Store actual similarity data from embeddings
embeddingsSimilarities: Map<string, number>;
}
interface ConfidenceMetrics {
overall: number; // 0-100: Combined confidence score
semanticRelevance: number; // How well tool description matches query (from embeddings)
taskSuitability: number; // AI-determined fitness for this specific task
methodologicalConsistency: number; // How well different analysis steps agree
uncertaintyFactors: string[]; // Specific reasons why this might not work
strengthIndicators: string[]; // Specific reasons why this is a good choice
overall: number;
semanticRelevance: number;
taskSuitability: number;
uncertaintyFactors: string[];
strengthIndicators: string[];
}
class ImprovedMicroTaskAIPipeline {
@ -102,10 +100,10 @@ class ImprovedMicroTaskAIPipeline {
};
private confidenceConfig: {
semanticWeight: number; // Weight for embeddings similarity
suitabilityWeight: number; // Weight for AI task fit evaluation
consistencyWeight: number; // Weight for cross-validation agreement
reliabilityWeight: number; // Weight for tool quality indicators
semanticWeight: number;
suitabilityWeight: number;
consistencyWeight: number;
reliabilityWeight: number;
minimumThreshold: number;
mediumThreshold: number;
highThreshold: number;
@ -143,10 +141,9 @@ class ImprovedMicroTaskAIPipeline {
retentionHours: parseInt(process.env.FORENSIC_AUDIT_RETENTION_HOURS || '72', 10)
};
// Updated confidence weights - more focused on AI evaluation
this.confidenceConfig = {
semanticWeight: parseFloat(process.env.CONFIDENCE_SEMANTIC_WEIGHT || '0.3'), // Embeddings similarity
suitabilityWeight: parseFloat(process.env.CONFIDENCE_SUITABILITY_WEIGHT || '0.7'), // AI task fit evaluation
semanticWeight: parseFloat(process.env.CONFIDENCE_SEMANTIC_WEIGHT || '0.3'),
suitabilityWeight: parseFloat(process.env.CONFIDENCE_SUITABILITY_WEIGHT || '0.7'),
consistencyWeight: 0,
reliabilityWeight: 0,
minimumThreshold: parseInt(process.env.CONFIDENCE_MINIMUM_THRESHOLD || '40', 10),
@ -235,7 +232,7 @@ class ImprovedMicroTaskAIPipeline {
const selectionRatio = result.selectedTools.length / candidateCount;
const hasReasoning = result.reasoning && result.reasoning.length > 50;
let confidence = 60; // Base confidence
let confidence = 60;
if (selectionRatio > 0.05 && selectionRatio < 0.3) confidence += 20;
else if (selectionRatio <= 0.05) confidence -= 10;
@ -386,7 +383,6 @@ class ImprovedMicroTaskAIPipeline {
let candidateConcepts: any[] = [];
let selectionMethod = 'unknown';
// Initialize embeddings similarities storage
context.embeddingsSimilarities = new Map<string, number>();
if (process.env.AI_EMBEDDINGS_ENABLED === 'true') {
@ -409,7 +405,6 @@ class ImprovedMicroTaskAIPipeline {
console.log(`[AI PIPELINE] Embeddings found ${similarItems.length} similar items`);
// Store actual similarity scores for confidence calculation
similarItems.forEach(item => {
context.embeddingsSimilarities.set(item.name, item.similarity);
});
@ -707,18 +702,14 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
limitations: string[] = []
): ConfidenceMetrics {
// 1. Semantic Relevance: Real embeddings similarity score
const rawSemanticRelevance = context.embeddingsSimilarities.has(tool.name) ?
context.embeddingsSimilarities.get(tool.name)! * 100 : 50;
// 2. Task Suitability: Enhanced with phase-awareness for workflow mode
let enhancedTaskSuitability = taskRelevance;
if (context.mode === 'workflow') {
// In workflow mode, boost score if tool is well-matched to its assigned phase
const toolSelection = context.selectedTools?.find(st => st.tool.name === tool.name);
if (toolSelection && tool.phases && tool.phases.includes(toolSelection.phase)) {
// Boost for phase alignment (but cap at 100)
const phaseBonus = Math.min(15, 100 - taskRelevance);
enhancedTaskSuitability = Math.min(100, taskRelevance + phaseBonus);
@ -726,7 +717,6 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
}
}
// Simple weighted combination - no artificial scaling
const overall = (
rawSemanticRelevance * this.confidenceConfig.semanticWeight +
enhancedTaskSuitability * this.confidenceConfig.suitabilityWeight
@ -747,7 +737,6 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
overall: Math.round(overall),
semanticRelevance: Math.round(rawSemanticRelevance),
taskSuitability: Math.round(enhancedTaskSuitability),
methodologicalConsistency: 0,
uncertaintyFactors,
strengthIndicators
};
@ -756,18 +745,15 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
private identifySpecificUncertaintyFactors(tool: any, context: AnalysisContext, limitations: string[], confidence: number): string[] {
const factors: string[] = [];
// Add AI-identified limitations first (most specific)
if (limitations && limitations.length > 0) {
factors.push(...limitations.slice(0, 2)); // Limit to top 2 to leave room for others
factors.push(...limitations.slice(0, 2));
}
// Low semantic similarity
const similarity = context.embeddingsSimilarities.get(tool.name) || 0.5;
if (similarity < 0.7) {
factors.push('Geringe semantische Ähnlichkeit zur Anfrage - Tool-Beschreibung passt möglicherweise nicht optimal');
}
// Skill level vs scenario complexity mismatch
if (tool.skillLevel === 'expert' && /schnell|rapid|triage|urgent|sofort/i.test(context.userQuery)) {
factors.push('Experten-Tool für zeitkritisches Szenario - Setup und Einarbeitung könnten zu lange dauern');
}
@ -776,35 +762,29 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
factors.push('Einsteiger-Tool für komplexe Analyse - könnte funktionale Limitierungen haben');
}
// Access and deployment concerns
if (tool.type === 'software' && !isToolHosted(tool) && tool.accessType === 'download') {
factors.push('Installation und Setup erforderlich');
}
// License restrictions
if (tool.license === 'Proprietary') {
factors.push('Kommerzielle Software - Lizenzkosten und rechtliche Beschränkungen zu beachten');
}
// Low overall confidence warning
if (confidence < 60) {
factors.push('Moderate Gesamtbewertung - alternative Ansätze sollten ebenfalls betrachtet werden');
}
return factors.slice(0, 4); // Limit to 4 most relevant factors
return factors.slice(0, 4);
}
// NEW: Identify specific strength indicators
private identifySpecificStrengthIndicators(tool: any, context: AnalysisContext, confidence: number): string[] {
const indicators: string[] = [];
// High semantic similarity
const similarity = context.embeddingsSimilarities.get(tool.name) || 0.5;
if (similarity >= 0.7) {
indicators.push('Sehr gute semantische Übereinstimmung mit Ihrer Anfrage');
}
// Quality indicators
if (tool.knowledgebase === true) {
indicators.push('Umfassende Dokumentation und Wissensbasis verfügbar');
}
@ -813,17 +793,15 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
indicators.push('Sofort verfügbar über gehostete Lösung - kein Setup erforderlich');
}
// Skill level match
if (tool.skillLevel === 'intermediate' || tool.skillLevel === 'advanced') {
indicators.push('Ausgewogenes Verhältnis zwischen Funktionalität und Benutzerfreundlichkeit');
}
// Method alignment
if (tool.type === 'method' && /methodik|vorgehen|prozess|ansatz/i.test(context.userQuery)) {
indicators.push('Methodischer Ansatz passt zu Ihrer prozeduralen Anfrage');
}
return indicators.slice(0, 4); // Limit to 4 most important indicators
return indicators.slice(0, 4);
}
private async analyzeScenario(context: AnalysisContext): Promise<MicroTaskResult> {
@ -902,11 +880,9 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
validSelections.forEach((sel: any) => {
const tool = phaseTools.find((t: any) => t.name === sel.toolName);
if (tool) {
// Ensure taskRelevance is a number
const taskRelevance = typeof sel.taskRelevance === 'number' ?
sel.taskRelevance : parseInt(String(sel.taskRelevance)) || 70;
// Derive priority automatically from score
const priority = this.derivePriorityFromScore(taskRelevance);
this.addToolToSelection(context, tool, phase.id, priority, sel.justification, taskRelevance, sel.limitations);
@ -967,7 +943,7 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
hasExplanation: !!evaluation.detailed_explanation,
hasImplementationApproach: !!evaluation.implementation_approach,
prosCount: evaluation.pros?.length || 0,
limitationsCount: evaluation.limitations?.length || 0, // ← Updated field name
limitationsCount: evaluation.limitations?.length || 0,
hasLimitations: Array.isArray(evaluation.limitations) && evaluation.limitations.length > 0
},
70,
@ -1101,7 +1077,7 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
const context: AnalysisContext = {
userQuery,
mode,
filteredData: {}, // Will be populated by getIntelligentCandidates
filteredData: {},
contextHistory: [],
maxContextLength: this.maxContextTokens,
currentContextLength: 0,
@ -1124,9 +1100,7 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
startTime,
{ auditEnabled: this.auditConfig.enabled, confidenceScoringEnabled: true }
);
// MICRO-TASK SEQUENCE WITH ENHANCED CONFIDENCE TRACKING
const analysisResult = await this.analyzeScenario(context);
if (analysisResult.success) completeTasks++; else failedTasks++;
await this.delay(this.microTaskDelay);
@ -1234,7 +1208,6 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
components: {
semantic: confidence.semanticRelevance,
suitability: confidence.taskSuitability,
consistency: confidence.methodologicalConsistency
}
},
confidence.overall,
@ -1286,7 +1259,7 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
detailed_explanation: st.tool.evaluation?.detailed_explanation || '',
implementation_approach: st.tool.evaluation?.implementation_approach || '',
pros: st.tool.evaluation?.pros || [],
cons: st.tool.evaluation?.limitations || [], // ← FIXED: Use limitations as cons for display
cons: st.tool.evaluation?.limitations || [],
alternatives: st.tool.evaluation?.alternatives || '',
confidence: confidence,
recommendationStrength: confidence.overall >= this.confidenceConfig.highThreshold ? 'strong' :

11
src/utils/embeddings.ts
View File

@ -31,7 +31,7 @@ interface SimilarityResult extends EmbeddingData {
class EmbeddingsService {
private embeddings: EmbeddingData[] = [];
private isInitialized = false;
private initializationPromise: Promise<void> | null = null; // ADD THIS LINE
private initializationPromise: Promise<void> | null = null;
private readonly embeddingsPath = path.join(process.cwd(), 'data', 'embeddings.json');
private readonly batchSize: number;
private readonly batchDelay: number;
@ -43,24 +43,19 @@ class EmbeddingsService {
this.batchDelay = parseInt(process.env.AI_EMBEDDINGS_BATCH_DELAY_MS || '1000', 10);
}
// REPLACE the existing initialize method with this:
async initialize(): Promise<void> {
// If initialization is already in progress, wait for it
if (this.initializationPromise) {
return this.initializationPromise;
}
// If already initialized, return immediately
if (this.isInitialized) {
return Promise.resolve();
}
// Start initialization and store the promise
this.initializationPromise = this.performInitialization();
return this.initializationPromise;
}
// ADD THIS NEW METHOD:
private async performInitialization(): Promise<void> {
if (!this.enabled) {
console.log('[EMBEDDINGS] Embeddings disabled, skipping initialization');
@ -70,13 +65,11 @@ class EmbeddingsService {
try {
console.log('[EMBEDDINGS] Initializing embeddings system...');
// Create data directory if it doesn't exist
await fs.mkdir(path.dirname(this.embeddingsPath), { recursive: true });
const toolsData = await getCompressedToolsDataForAI();
const currentDataHash = this.hashData(toolsData);
// Try to load existing embeddings
const existingEmbeddings = await this.loadEmbeddings();
if (existingEmbeddings && existingEmbeddings.version === currentDataHash) {
@ -336,12 +329,10 @@ class EmbeddingsService {
// Global instance
const embeddingsService = new EmbeddingsService();
export { embeddingsService, type EmbeddingData, type SimilarityResult };
// Auto-initialize on import in server environment
if (typeof window === 'undefined' && process.env.NODE_ENV !== 'test') {
embeddingsService.initialize().catch(error => {
console.error('[EMBEDDINGS] Auto-initialization failed:', error);

3
src/utils/rateLimitedQueue.ts
View File

@ -96,7 +96,6 @@ class RateLimitedQueue {
this.tasks.push(queuedTask);
// Kick the processor soon.
setTimeout(() => {
this.processQueue();
}, 100);
@ -170,7 +169,7 @@ class RateLimitedQueue {
.filter((t) => t.status === "queued")
.sort((a, b) => a.addedAt - b.addedAt)[0];
if (!nextTask) break; // No more work
if (!nextTask) break;
nextTask.status = "processing";
nextTask.startedAt = Date.now();