main #11
16
.env.example
16
.env.example
@ -80,6 +80,14 @@ AI_SIMILARITY_THRESHOLD=0.3
|
||||
AI_EMBEDDING_SELECTION_LIMIT=30
|
||||
AI_EMBEDDING_CONCEPTS_LIMIT=15
|
||||
|
||||
# === METHOD/TOOL BALANCE CONFIGURATION ===
|
||||
# Controls the ratio of methods vs software tools sent to AI
|
||||
# Methods = procedural guidance, best practices, workflows
|
||||
# Software = actual tools and applications
|
||||
# Values should sum to less than 1.0 (remainder is buffer)
|
||||
AI_METHOD_SELECTION_RATIO=0.4 # 40% methods (increase for more procedural guidance)
|
||||
AI_SOFTWARE_SELECTION_RATIO=0.5 # 50% software tools (increase for more tool recommendations)
|
||||
|
||||
# AI selection limits
|
||||
AI_MAX_SELECTED_ITEMS=25
|
||||
AI_MAX_TOOLS_TO_ANALYZE=20
|
||||
@ -155,6 +163,14 @@ CONFIDENCE_HIGH_THRESHOLD=80
|
||||
# AI_MAX_TOOLS_TO_ANALYZE=10
|
||||
# AI_EMBEDDINGS_ENABLED=false
|
||||
|
||||
# 🔬 METHOD-FOCUSED (more procedural guidance, less tools):
|
||||
# AI_METHOD_SELECTION_RATIO=0.6
|
||||
# AI_SOFTWARE_SELECTION_RATIO=0.3
|
||||
|
||||
# 🛠️ TOOL-FOCUSED (more software recommendations, less methods):
|
||||
# AI_METHOD_SELECTION_RATIO=0.2
|
||||
# AI_SOFTWARE_SELECTION_RATIO=0.7
|
||||
|
||||
# ============================================================================
|
||||
# 🌐 AI SERVICE EXAMPLES
|
||||
# ============================================================================
|
||||
|
||||
@ -1,212 +1,207 @@
|
||||
// src/config/prompts.ts
|
||||
// src/config/prompts.ts - Centralized German prompts for AI pipeline
|
||||
|
||||
export const AI_PROMPTS = {
|
||||
|
||||
toolSelection: (mode: string, userQuery: string, selectionMethod: string, maxSelectedItems: number) => {
|
||||
const modeInstruction = mode === 'workflow'
|
||||
? 'Der Benutzer möchte einen UMFASSENDEN WORKFLOW über mehrere Phasen. Wählen Sie 15–25 Elemente, die den Zyklus Datensammlung → Auswertung → Analyse → Bericht abdecken.'
|
||||
: 'Der Benutzer möchte SPEZIFISCHE LÖSUNGEN. Wählen Sie 4–10 Elemente, die das Problem direkt adressieren.';
|
||||
? 'Workflow mit 15-25 Items über alle Phasen. PFLICHT: Mindestens 40% Methoden, Rest Tools/Konzepte.'
|
||||
: 'Spezifische Lösung mit 4-10 Items. PFLICHT: Mindestens 30% Methoden wenn verfügbar.';
|
||||
|
||||
return `Sie sind DFIR-Experte mit Zugriff auf eine vollständige Datenbank aus Methoden (type: "method") und Tools (type: "software"/"os"). Ihre Aufgabe ist es, die relevantesten Elemente für die Anfrage auszuwählen.
|
||||
return `Du bist ein DFIR-Experte. Wähle die BESTEN Items aus dem vorgefilterten Set.
|
||||
|
||||
AUSWAHLMETHODE: ${selectionMethod}
|
||||
${selectionMethod === 'embeddings_candidates' ?
|
||||
'Die angezeigten Kandidaten wurden bereits semantisch vorgefiltert. Wählen Sie daraus die BESTEN.' :
|
||||
'Sie sehen die vollständige Datenbasis. Wählen Sie die relevantesten Elemente.'}
|
||||
|
||||
WICHTIG – MISCHUNG ERZWINGEN:
|
||||
- Der finale Vorschlag MUSS eine Mischung aus Methoden UND Tools enthalten (sofern verfügbar).
|
||||
- Ziel: grob 40–60 % Methoden und 40–60 % Tools. Wenn eine Kategorie knapp ist, wählen Sie so ausgewogen wie möglich. Begründen Sie Abweichungen im reasoning.
|
||||
- Bezeichnen Sie Elemente konsequent als "Methode" oder "Tool" je nach Typ.
|
||||
'✓ Semantisch relevante Items bereits vorgefiltert\n✓ Wähle die BESTEN für die konkrete Aufgabe' :
|
||||
'✓ Vollständige Datenbank verfügbar\n✓ Wähle die relevantesten Items'}
|
||||
|
||||
${modeInstruction}
|
||||
|
||||
BENUTZER-ANFRAGE: "${userQuery}"
|
||||
ANFRAGE: "${userQuery}"
|
||||
|
||||
KRITISCHE AUSWAHLPRINZIPIEN:
|
||||
1) KONTEXT > POPULARITÄT: Nicht automatisch zu populären Tools greifen. Wählen Sie strikt szenariobezogen.
|
||||
2) METHODIK vs. SOFTWARE:
|
||||
- Dringend/Triage → Methoden für schnelle Reaktion priorisieren.
|
||||
- Zeitkritisch → Triage-Methoden vor Tiefenanalyse-Tools.
|
||||
- Umfassende Analyse → dann spezialisierte Tools ergänzen.
|
||||
3) SPEZIFITÄT:
|
||||
- ICS/SCADA → spezialisierte ICS-Methoden/-Tools statt generischer Netzwerk-Tools.
|
||||
- Mobile (Android/iOS) → mobile-spezifische Methoden/Tools.
|
||||
- Speicheranalyse dringend → schnelle Memory-Methoden/Tools vor Vollanalyse.
|
||||
4) LESEN SIE DIE VOLLBESCHREIBUNG UND METADATEN (Tags, Plattformen, Phasen, Lizenz, Access).
|
||||
VERFÜGBARE ITEM-TYPEN:
|
||||
- TOOLS (type: "software"/"method") → praktische Anwendungen und Vorgehensweisen
|
||||
- KONZEPTE (type: "concept") → theoretisches Wissen und Methodiken
|
||||
|
||||
UI-KÜRZE:
|
||||
- Denken Sie mit: Für das UI sind kurze, prägnante Beschreibungen nötig. Bevorzugen Sie Elemente, die sich knapp zusammenfassen lassen. Formulieren Sie Begründungen fokussiert und ohne Floskeln.
|
||||
AUSWAHLSTRATEGIE:
|
||||
1. **ERSTE PRIORITÄT: Relevanz zur Anfrage**
|
||||
- Direkt anwendbar auf das Problem
|
||||
- Löst die Kernherausforderung
|
||||
|
||||
Wählen Sie die relevantesten Elemente (max ${maxSelectedItems} gesamt).
|
||||
2. **ZWEITE PRIORITÄT: Ausgewogene Mischung**
|
||||
- Tools/Methoden für praktische Umsetzung → selectedTools
|
||||
- Konzepte für methodisches Verständnis → selectedConcepts
|
||||
- WICHTIG: Auch Konzepte auswählen, nicht nur Tools!
|
||||
|
||||
ANTWORTFORMAT (JSON, SCHEMA UNVERÄNDERT):
|
||||
3. **QUALITÄT > QUANTITÄT**
|
||||
- Lieber weniger perfekte Items als viele mittelmäßige
|
||||
- Jedes Item muss begründbar sein
|
||||
|
||||
AUSWAHLREGELN:
|
||||
- Wähle ${mode === 'workflow' ? '15-25' : '4-10'} Items total, max ${maxSelectedItems}
|
||||
- BEIDE Arrays füllen: selectedTools UND selectedConcepts
|
||||
- Mindestens 1-2 Konzepte auswählen für methodische Fundierung
|
||||
- Tools: 40% Methoden (type="method"), Rest Software (type="software")
|
||||
|
||||
ANTWORT AUSSCHLIESSLICH IM JSON-FORMAT:
|
||||
{
|
||||
"selectedTools": ["Name 1", "Name 2", ...],
|
||||
"selectedConcepts": ["Konzept 1", "Konzept 2", ...],
|
||||
"reasoning": "Begründen Sie die Auswahl, die erzwungene Mischung Methode/Tool und warum bestimmte populäre Elemente nicht geeignet sind. Halten Sie die Sprache präzise und UI-tauglich."
|
||||
"selectedTools": ["ToolName1", "MethodName1", ...],
|
||||
"selectedConcepts": ["ConceptName1", "ConceptName2", ...],
|
||||
"reasoning": "Kurze Begründung mit Erwähnung der Tool/Konzept-Balance"
|
||||
}`;
|
||||
},
|
||||
|
||||
scenarioAnalysis: (isWorkflow: boolean, userQuery: string) => {
|
||||
const analysisType = isWorkflow ? 'Untersuchungsszenario' : 'technische Problemstellung';
|
||||
const considerations = isWorkflow ?
|
||||
`- Angriffsvektoren (MITRE ATT&CK) und Bedrohungsmodell
|
||||
- Betroffene Systeme/Assets und Kritikalität
|
||||
- Zeitdruck, Beweiserhalt, Chain of Custody
|
||||
- Relevante Artefakte und Datenquellen` :
|
||||
`- Konkrete forensische Hürden
|
||||
- Verfügbare Datenquellen und Integrität
|
||||
- Anforderungen für rechtssichere Auswertung`;
|
||||
const analysisType = isWorkflow ? 'Szenario' : 'Problem';
|
||||
const focus = isWorkflow ?
|
||||
'Angriffsvektoren, betroffene Systeme, Zeitkritikalität' :
|
||||
'Kernherausforderung, verfügbare Daten, methodische Anforderungen';
|
||||
|
||||
return `Analysieren Sie das ${analysisType} prägnant in einem kurzen Fließtext (max. 220 Wörter).
|
||||
return `DFIR-Experte: Analysiere das ${analysisType}.
|
||||
|
||||
${isWorkflow ? 'SZENARIO' : 'PROBLEM'}: "${userQuery}"
|
||||
|
||||
Berücksichtigen:
|
||||
${considerations}
|
||||
Fokus: ${focus}
|
||||
|
||||
Stil: Deutsch, ohne Listen/Markdown; klare, knappe Formulierungen, aber nicht telegrafisch (vollständige Sätze).`;
|
||||
Antwort: Fließtext ohne Listen, max 100 Wörter.`;
|
||||
},
|
||||
|
||||
investigationApproach: (isWorkflow: boolean, userQuery: string) => {
|
||||
const approachType = isWorkflow ? 'Untersuchungsansatz' : 'Lösungsansatz';
|
||||
const considerations = isWorkflow ?
|
||||
`- Triage-Prioritäten nach forensischer Dringlichkeit
|
||||
- Phasenabfolge: Datensammlung → Auswertung → Analyse → Bericht
|
||||
- Kontaminationsvermeidung/Isolierung
|
||||
- Objektivität und Nachvollziehbarkeit` :
|
||||
`- Methodenwahl nach wissenschaftlichen Kriterien
|
||||
- Validierung/Verifizierung der Vorgehensweise
|
||||
- Integration in bestehende DFIR-Workflows
|
||||
- Reproduzierbarkeit und Transparenz`;
|
||||
const focus = isWorkflow ?
|
||||
'Triage-Prioritäten, Phasenabfolge, Kontaminationsvermeidung' :
|
||||
'Methodenauswahl, Validierung, Integration';
|
||||
|
||||
return `Formulieren Sie einen knappen ${approachType} (max. 220 Wörter), der explizit eine Mischung aus Methoden und Tools vorsieht.
|
||||
return `Entwickle einen ${approachType}.
|
||||
|
||||
${isWorkflow ? 'SZENARIO' : 'PROBLEM'}: "${userQuery}"
|
||||
|
||||
Berücksichtigen:
|
||||
${considerations}
|
||||
Fokus: ${focus}
|
||||
|
||||
Stil: Deutsch, ohne Listen/Markdown, UI-tauglich und fokussiert.`;
|
||||
Antwort: Fließtext ohne Listen, max 100 Wörter.`;
|
||||
},
|
||||
|
||||
criticalConsiderations: (isWorkflow: boolean, userQuery: string) => {
|
||||
const considerationType = isWorkflow ? 'kritische forensische Überlegungen' : 'methodische Voraussetzungen';
|
||||
const aspects = isWorkflow ?
|
||||
`- Beweissicherung vs. Gründlichkeit (Zeitdruck)
|
||||
- Chain of Custody und rechtliche Verwertbarkeit
|
||||
- Transparenz, Reproduzierbarkeit, Dokumentationspflichten` :
|
||||
`- Validierung/Nachvollziehbarkeit der Methode/Tools
|
||||
- Risiken für False Positives/Negatives
|
||||
- Qualifikationsanforderungen und Reporting-Standards`;
|
||||
const focus = isWorkflow ?
|
||||
'Beweissicherung vs. Gründlichkeit, Chain of Custody' :
|
||||
'Tool-Validierung, False Positives/Negatives, Qualifikationen';
|
||||
|
||||
return `Identifizieren Sie ${considerationType} in einem kurzen Fließtext (max. 220 Wörter) und achten Sie auf UI-Kürze.
|
||||
return `Identifiziere kritische Überlegungen.
|
||||
|
||||
${isWorkflow ? 'SZENARIO' : 'PROBLEM'}: "${userQuery}"
|
||||
|
||||
Aspekte:
|
||||
${aspects}
|
||||
Fokus: ${focus}
|
||||
|
||||
Stil: Deutsch, ohne Listen/Markdown.`;
|
||||
Antwort: Fließtext ohne Listen, max 100 Wörter.`;
|
||||
},
|
||||
|
||||
phaseToolSelection: (userQuery: string, phase: any, phaseTools: any[]) => {
|
||||
// In der Liste bewusst kurze Ausschnitte und klare Typ-Bezeichnung (Methode/Tool).
|
||||
const items = phaseTools.map((tool: any, index: number) => {
|
||||
const typ = tool.type === 'method' ? 'Methode' : 'Tool';
|
||||
const desc = (tool.description || '').replace(/\s+/g, ' ').trim().slice(0, 320);
|
||||
return `${index + 1}. [${typ}] ${tool.name}: ${desc}…\n - Plattformen: ${tool.platforms?.join(', ') || 'N/A'}\n - Skill Level: ${tool.skillLevel}\n - Tags: ${tool.tags?.join(', ') || 'N/A'}`;
|
||||
}).join('\n\n');
|
||||
|
||||
return `Wählen Sie 2–3 Elemente für die Phase "${phase.name}" und bewerten Sie diese VERGLEICHEND. Erzwingen Sie eine Mischung aus mindestens einer Methode und einem Tool, sofern verfügbar.
|
||||
const methods = phaseTools.filter(t => t.type === 'method');
|
||||
const tools = phaseTools.filter(t => t.type === 'software');
|
||||
|
||||
if (phaseTools.length === 0) {
|
||||
return `Keine Tools für Phase "${phase.name}" verfügbar. Antworte mit leerem Array: []`;
|
||||
}
|
||||
|
||||
return `Du bist ein DFIR-Experte. Wähle die 2-3 BESTEN Items für Phase "${phase.name}".
|
||||
|
||||
SZENARIO: "${userQuery}"
|
||||
PHASE: ${phase.name} – ${phase.description || 'Forensische Untersuchungsphase'}
|
||||
PHASE: ${phase.name} - ${phase.description || ''}
|
||||
|
||||
VERFÜGBARE ELEMENTE FÜR ${phase.name.toUpperCase()}:
|
||||
${items}
|
||||
VERFÜGBARE ITEMS (bereits von KI vorausgewählt):
|
||||
${methods.length > 0 ? `
|
||||
METHODEN (${methods.length}):
|
||||
${methods.map((method: any) =>
|
||||
`- ${method.name}
|
||||
Typ: ${method.type}
|
||||
Beschreibung: ${method.description.slice(0, 150)}...
|
||||
Domains: ${method.domains?.join(', ') || 'N/A'}
|
||||
Skill Level: ${method.skillLevel}`
|
||||
).join('\n\n')}
|
||||
` : 'Keine Methoden verfügbar'}
|
||||
|
||||
BEWERTUNGSKRITERIEN:
|
||||
- Eignung für das konkrete Szenario und GENAU diese Phase
|
||||
- Vergleich zu den Alternativen in der Liste
|
||||
- Praktikabilität (Zeit, Setup, Reproduzierbarkeit)
|
||||
${tools.length > 0 ? `
|
||||
SOFTWARE TOOLS (${tools.length}):
|
||||
${tools.map((tool: any) =>
|
||||
`- ${tool.name}
|
||||
Typ: ${tool.type}
|
||||
Beschreibung: ${tool.description.slice(0, 150)}...
|
||||
Plattformen: ${tool.platforms?.join(', ') || 'N/A'}
|
||||
Skill Level: ${tool.skillLevel}`
|
||||
).join('\n\n')}
|
||||
` : 'Keine Software-Tools verfügbar'}
|
||||
|
||||
UI-KÜRZE:
|
||||
- Beginnen Sie in "justification" mit einer 2–4-Satz Kurzfassung (≈300–360 Zeichen), danach 2–3 kurze, präzise Sätze mit Details.
|
||||
- Bezeichnen Sie Elemente als "Methode" oder "Tool".
|
||||
AUSWAHLREGELN FÜR PHASE "${phase.name}":
|
||||
1. Wähle die 2-3 BESTEN Items für diese spezifische Phase
|
||||
2. Priorisiere Items, die DIREKT für "${phase.name}" relevant sind
|
||||
3. Mindestens 1 Methode wenn verfügbar, Rest Software-Tools
|
||||
4. Begründe WARUM jedes Item für diese Phase optimal ist
|
||||
|
||||
ANTWORT NUR ALS JSON (Schema UNVERÄNDERT):
|
||||
WICHTIG: Verwende EXAKT die Namen wie oben aufgelistet (ohne Präfixe wie M1./T2.)!
|
||||
|
||||
ANTWORT AUSSCHLIESSLICH IM JSON-FORMAT OHNE JEGLICHEN TEXT AUSSERHALB:
|
||||
[
|
||||
{
|
||||
"toolName": "Exakter Methoden/Tool-Name",
|
||||
"taskRelevance": 0–100,
|
||||
"justification": "Kurzfassung zuerst (≈300–360 Zeichen). Danach knappe, vergleichende Begründung, warum dieses Element in dieser Phase besser geeignet ist.",
|
||||
"limitations": ["Konkrete Einschränkung 1", "Einschränkung 2"]
|
||||
"toolName": "Exakter Name aus der Liste oben",
|
||||
"taskRelevance": 85,
|
||||
"justification": "Spezifische Begründung warum optimal für ${phase.name}",
|
||||
"limitations": ["Mögliche Einschränkung für diese Phase"]
|
||||
}
|
||||
]
|
||||
|
||||
Wählen Sie nur die 2–3 BESTEN Elemente und achten Sie auf die Mischung Methode+Tool.`;
|
||||
]`;
|
||||
},
|
||||
|
||||
toolEvaluation: (userQuery: string, tool: any, rank: number, taskRelevance: number) => {
|
||||
const typ = tool.type === 'method' ? 'Methode' : 'Tool';
|
||||
return `Sie sind DFIR-Experte. Erklären Sie DETAILLIERT die Anwendung dieses bereits bewerteten Elements.
|
||||
const itemType = tool.type === 'method' ? 'Methode' : 'Tool';
|
||||
|
||||
return `Erkläre die Anwendung dieser/dieses ${itemType}.
|
||||
|
||||
PROBLEM: "${userQuery}"
|
||||
ELEMENT: ${typ} ${tool.name} (Aufgaben-Eignung ${taskRelevance}%)
|
||||
BESCHREIBUNG: ${(tool.description || '').replace(/\s+/g, ' ').trim()}
|
||||
${itemType.toUpperCase()}: ${tool.name} (${taskRelevance}% Eignung)
|
||||
TYP: ${tool.type}
|
||||
|
||||
ANTWORT AUSSCHLIESSLICH ALS JSON (Schema UNVERÄNDERT):
|
||||
Bereits als Rang ${rank} bewertet.
|
||||
|
||||
ANTWORT AUSSCHLIESSLICH IM JSON-FORMAT OHNE JEGLICHEN TEXT AUSSERHALB DER JSON-STRUKTUR:
|
||||
{
|
||||
"detailed_explanation": "Beginnen Sie mit einer prägnanten Kurzfassung (≈300–360 Zeichen) für das UI: 'Kurz: …'. Danach eine verdichtete Hauptpassage (≈80–140 Wörter) mit präziser, technisch korrekter Erläuterung, warum und wie diese ${typ.toLowerCase()} für dieses Problem eingesetzt wird.",
|
||||
"implementation_approach": "Konkrete, reproduzierbare Schrittfolge. Kurze Schritte, kein Ausschweifen.",
|
||||
"pros": ["Spezifischer Vorteil 1", "Vorteil 2"],
|
||||
"limitations": ["Konkrete Einschränkung 1", "Einschränkung 2"],
|
||||
"alternatives": "Knappe Alternativen, falls dieses Element nicht verfügbar ist (Methoden und Tools anführen)."
|
||||
}
|
||||
|
||||
WICHTIG:
|
||||
- Keine erneute Bewertung – nur Erklärung/Anwendung.
|
||||
- Behalten Sie die korrekte Bezeichnung (Methode/Tool) bei.
|
||||
- UI-Kürze: Kurzfassung am Anfang der detailed_explanation.`;
|
||||
"detailed_explanation": "Warum und wie einsetzen",
|
||||
"implementation_approach": "Konkrete Schritte",
|
||||
"pros": ["Vorteil 1", "Vorteil 2"],
|
||||
"limitations": ["Einschränkung 1"],
|
||||
"alternatives": "Alternative Ansätze"
|
||||
}`;
|
||||
},
|
||||
|
||||
backgroundKnowledgeSelection: (userQuery: string, mode: string, selectedToolNames: string[], availableConcepts: any[]) => {
|
||||
const conceptsList = availableConcepts.slice(0, 15).map((concept: any) => `- ${concept.name}: ${(concept.description || '').replace(/\s+/g, ' ').trim().slice(0, 80)}…`).join('\n');
|
||||
return `Wählen Sie 2–4 forensische Konzepte, die für die Anwendung der empfohlenen Elemente (Methoden + Tools) wesentlich sind.
|
||||
return `Wähle 2-4 relevante Konzepte.
|
||||
|
||||
${mode === 'workflow' ? 'SZENARIO' : 'PROBLEM'}: "${userQuery}"
|
||||
EMPFOHLENE ELEMENTE: ${selectedToolNames.join(', ')}
|
||||
AUSGEWÄHLTE TOOLS: ${selectedToolNames.join(', ')}
|
||||
|
||||
VERFÜGBARE KONZEPTE:
|
||||
${conceptsList}
|
||||
VERFÜGBARE KONZEPTE (${availableConcepts.length} KI-kuratiert):
|
||||
${availableConcepts.map((c: any) =>
|
||||
`- ${c.name}: ${c.description}...`
|
||||
).join('\n')}
|
||||
|
||||
ANTWORT NUR ALS JSON (Schema UNVERÄNDERT):
|
||||
ANTWORT AUSSCHLIESSLICH IM JSON-FORMAT OHNE JEGLICHEN TEXT AUSSERHALB DER JSON-STRUKTUR:
|
||||
[
|
||||
{
|
||||
"conceptName": "Exakter Konzept-Name",
|
||||
"relevance": "Knappe Begründung, warum dieses Konzept für das Verständnis der Methodik/Tools kritisch ist (UI-tauglich)."
|
||||
"conceptName": "Name",
|
||||
"relevance": "Warum kritisch für Methodik"
|
||||
}
|
||||
]`;
|
||||
},
|
||||
|
||||
finalRecommendations: (isWorkflow: boolean, userQuery: string, selectedToolNames: string[]) => {
|
||||
if (isWorkflow) {
|
||||
return `Erstellen Sie einen knappen, methodisch korrekten WORKFLOW-Fließtext (max. 220 Wörter), der EXPLIZIT eine Mischung aus Methoden und Tools nutzt.
|
||||
const focus = isWorkflow ?
|
||||
'Workflow-Schritte, Best Practices, Objektivität' :
|
||||
'Methodische Überlegungen, Validierung, Qualitätssicherung';
|
||||
|
||||
SZENARIO: "${userQuery}"
|
||||
AUSGEWÄHLTE ELEMENTE: ${selectedToolNames.join(', ') || 'Keine Auswahl'}
|
||||
return `Erstelle ${isWorkflow ? 'Workflow-Empfehlung' : 'methodische Überlegungen'}.
|
||||
|
||||
Vorgaben: klare Phasen (Datensammlung → Auswertung → Analyse → Bericht), Beweisführung/Chain of Custody, Reproduzierbarkeit. Sprache präzise und UI-tauglich. Keine Listen/Markdown.`;
|
||||
}
|
||||
${isWorkflow ? 'SZENARIO' : 'PROBLEM'}: "${userQuery}"
|
||||
AUSGEWÄHLT: ${selectedToolNames.join(', ')}${selectedToolNames.length > 5 ? '...' : ''}
|
||||
|
||||
return `Formulieren Sie knappe, kritische ÜBERLEGUNGEN (max. 200 Wörter) zur korrekten Anwendung der empfohlenen Elemente (Methoden + Tools).
|
||||
Fokus: ${focus}
|
||||
|
||||
PROBLEM: "${userQuery}"
|
||||
EMPFOHLENE ELEMENTE: ${selectedToolNames.join(', ') || 'Keine Auswahl'}
|
||||
|
||||
Fokus: Validierung/Qualitätssicherung, Risiken (FP/FN), Dokumentation/Reporting. Stil deutsch, ohne Listen/Markdown, UI-tauglich.`;
|
||||
Antwort: Fließtext ohne Listen, max ${isWorkflow ? '100' : '80'} Wörter.`;
|
||||
}
|
||||
} as const;
|
||||
|
||||
@ -231,4 +226,4 @@ export function getPrompt(promptKey: keyof typeof AI_PROMPTS, ...args: any[]): s
|
||||
console.error(`[PROMPTS] Error generating prompt ${promptKey}:`, error);
|
||||
return 'Error: Failed to generate prompt';
|
||||
}
|
||||
}
|
||||
}
|
||||
@ -67,6 +67,8 @@ interface AnalysisContext {
|
||||
auditTrail: AuditEntry[];
|
||||
|
||||
embeddingsSimilarities: Map<string, number>;
|
||||
aiSelectedTools?: any[];
|
||||
aiSelectedConcepts?: any[];
|
||||
}
|
||||
|
||||
interface ConfidenceMetrics {
|
||||
@ -93,6 +95,9 @@ class ImprovedMicroTaskAIPipeline {
|
||||
private embeddingsMinTools: number;
|
||||
private embeddingsMaxReductionRatio: number;
|
||||
|
||||
private methodSelectionRatio: number;
|
||||
private softwareSelectionRatio: number;
|
||||
|
||||
private maxContextTokens: number;
|
||||
private maxPromptTokens: number;
|
||||
|
||||
@ -127,12 +132,15 @@ class ImprovedMicroTaskAIPipeline {
|
||||
this.embeddingSelectionLimit = parseInt(process.env.AI_EMBEDDING_SELECTION_LIMIT || '30', 10);
|
||||
this.embeddingConceptsLimit = parseInt(process.env.AI_EMBEDDING_CONCEPTS_LIMIT || '15', 10);
|
||||
|
||||
this.noEmbeddingsToolLimit = parseInt(process.env.AI_NO_EMBEDDINGS_TOOL_LIMIT || '0', 10);
|
||||
this.noEmbeddingsConceptLimit = parseInt(process.env.AI_NO_EMBEDDINGS_CONCEPT_LIMIT || '0', 10);
|
||||
this.noEmbeddingsToolLimit = parseInt(process.env.AI_NO_EMBEDDINGS_TOOL_LIMIT || '25', 10);
|
||||
this.noEmbeddingsConceptLimit = parseInt(process.env.AI_NO_EMBEDDINGS_CONCEPT_LIMIT || '10', 10);
|
||||
|
||||
this.embeddingsMinTools = parseInt(process.env.AI_EMBEDDINGS_MIN_TOOLS || '8', 10);
|
||||
this.embeddingsMaxReductionRatio = parseFloat(process.env.AI_EMBEDDINGS_MAX_REDUCTION_RATIO || '0.75');
|
||||
|
||||
this.methodSelectionRatio = parseFloat(process.env.AI_METHOD_SELECTION_RATIO || '0.4');
|
||||
this.softwareSelectionRatio = parseFloat(process.env.AI_SOFTWARE_SELECTION_RATIO || '0.5');
|
||||
|
||||
this.maxContextTokens = parseInt(process.env.AI_MAX_CONTEXT_TOKENS || '4000', 10);
|
||||
this.maxPromptTokens = parseInt(process.env.AI_MAX_PROMPT_TOKENS || '1500', 10);
|
||||
|
||||
@ -142,6 +150,7 @@ class ImprovedMicroTaskAIPipeline {
|
||||
};
|
||||
|
||||
console.log('[AI PIPELINE] Audit trail enabled:', this.auditConfig.enabled);
|
||||
console.log('[AI PIPELINE] Method/Software balance:', `${(this.methodSelectionRatio * 100).toFixed(0)}%/${(this.softwareSelectionRatio * 100).toFixed(0)}%`);
|
||||
|
||||
this.confidenceConfig = {
|
||||
semanticWeight: parseFloat(process.env.CONFIDENCE_SEMANTIC_WEIGHT || '0.3'),
|
||||
@ -163,7 +172,7 @@ class ImprovedMicroTaskAIPipeline {
|
||||
FORENSIC_AUDIT_DETAIL_LEVEL: process.env.FORENSIC_AUDIT_DETAIL_LEVEL,
|
||||
NODE_ENV: process.env.NODE_ENV,
|
||||
allEnvKeys: Object.keys(process.env).filter(k => k.includes('AUDIT')),
|
||||
dotenvLoaded: !!process.env.PUBLIC_BASE_URL // Proxy for "dotenv worked"
|
||||
dotenvLoaded: !!process.env.PUBLIC_BASE_URL
|
||||
});
|
||||
|
||||
console.log('[AI PIPELINE] Final audit config:', this.auditConfig);
|
||||
@ -260,11 +269,40 @@ class ImprovedMicroTaskAIPipeline {
|
||||
|
||||
private safeParseJSON(jsonString: string, fallback: any = null): any {
|
||||
try {
|
||||
let cleaned = jsonString
|
||||
.replace(/^```json\s*/i, '')
|
||||
.replace(/\s*```\s*$/g, '')
|
||||
.trim();
|
||||
// First, try to extract JSON block from markdown or mixed content
|
||||
let cleaned = jsonString.trim();
|
||||
|
||||
// Look for JSON block patterns (most specific first)
|
||||
const jsonBlockPatterns = [
|
||||
/```json\s*([\s\S]*?)\s*```/i, // ```json ... ```
|
||||
/```\s*([\s\S]*?)\s*```/i, // ``` ... ```
|
||||
/\{[\s\S]*\}/, // Anything that looks like JSON object
|
||||
];
|
||||
|
||||
let jsonMatch: RegExpMatchArray | null = null;
|
||||
for (const pattern of jsonBlockPatterns) {
|
||||
jsonMatch = cleaned.match(pattern);
|
||||
if (jsonMatch) {
|
||||
cleaned = jsonMatch[1] || jsonMatch[0];
|
||||
console.log('[AI PIPELINE] Extracted JSON block using pattern:', pattern.source);
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
// If no pattern matched, try to find JSON-like content manually
|
||||
if (!jsonMatch) {
|
||||
const jsonStart = cleaned.indexOf('{');
|
||||
const jsonEnd = cleaned.lastIndexOf('}');
|
||||
if (jsonStart !== -1 && jsonEnd !== -1 && jsonEnd > jsonStart) {
|
||||
cleaned = cleaned.substring(jsonStart, jsonEnd + 1);
|
||||
console.log('[AI PIPELINE] Manually extracted JSON from position', jsonStart, 'to', jsonEnd);
|
||||
}
|
||||
}
|
||||
|
||||
// Clean up the extracted content
|
||||
cleaned = cleaned.trim();
|
||||
|
||||
// Handle truncated JSON by finding the last complete structure
|
||||
if (!cleaned.endsWith('}') && !cleaned.endsWith(']')) {
|
||||
console.warn('[AI PIPELINE] JSON appears truncated, attempting recovery...');
|
||||
|
||||
@ -305,9 +343,10 @@ class ImprovedMicroTaskAIPipeline {
|
||||
}
|
||||
|
||||
if (lastCompleteStructure) {
|
||||
console.log('[AI PIPELINE] Attempting to parse recovered JSON structure...');
|
||||
console.log('[AI PIPELINE] Using recovered JSON structure');
|
||||
cleaned = lastCompleteStructure;
|
||||
} else {
|
||||
// Try to close unclosed braces/brackets
|
||||
if (braceCount > 0) {
|
||||
cleaned += '}';
|
||||
console.log('[AI PIPELINE] Added closing brace to truncated JSON');
|
||||
@ -321,40 +360,82 @@ class ImprovedMicroTaskAIPipeline {
|
||||
|
||||
const parsed = JSON.parse(cleaned);
|
||||
|
||||
// Ensure the structure has the expected fields
|
||||
if (parsed && typeof parsed === 'object') {
|
||||
if (parsed.selectedTools === undefined) parsed.selectedTools = [];
|
||||
if (parsed.selectedConcepts === undefined) parsed.selectedConcepts = [];
|
||||
|
||||
if (!Array.isArray(parsed.selectedTools)) parsed.selectedTools = [];
|
||||
if (!Array.isArray(parsed.selectedConcepts)) parsed.selectedConcepts = [];
|
||||
|
||||
console.log(`[AI PIPELINE] Successfully parsed JSON: ${parsed.selectedTools.length} tools, ${parsed.selectedConcepts.length} concepts`);
|
||||
}
|
||||
|
||||
return parsed;
|
||||
|
||||
} catch (error) {
|
||||
console.warn('[AI PIPELINE] JSON parsing failed:', error.message);
|
||||
console.warn('[AI PIPELINE] Raw content (first 300 chars):', jsonString.slice(0, 300));
|
||||
console.warn('[AI PIPELINE] Raw content (last 300 chars):', jsonString.slice(-300));
|
||||
|
||||
if (jsonString.includes('selectedTools')) {
|
||||
const toolMatches = jsonString.match(/"([^"]+)"/g);
|
||||
if (toolMatches && toolMatches.length > 0) {
|
||||
console.log('[AI PIPELINE] Attempting partial recovery from broken JSON...');
|
||||
const possibleTools = toolMatches
|
||||
.map(match => match.replace(/"/g, ''))
|
||||
.filter(name => name.length > 2 && !['selectedTools', 'selectedConcepts', 'reasoning'].includes(name))
|
||||
.slice(0, 15);
|
||||
|
||||
if (possibleTools.length > 0) {
|
||||
console.log(`[AI PIPELINE] Recovered ${possibleTools.length} possible tool names from broken JSON`);
|
||||
return {
|
||||
selectedTools: possibleTools,
|
||||
selectedConcepts: [],
|
||||
reasoning: 'Recovered from truncated response'
|
||||
};
|
||||
// Enhanced recovery mechanism that preserves tool/concept distinction
|
||||
if (jsonString.includes('selectedTools') || jsonString.includes('selectedConcepts')) {
|
||||
console.log('[AI PIPELINE] Attempting enhanced recovery from broken JSON...');
|
||||
|
||||
// Try to extract tool and concept arrays separately
|
||||
const selectedTools: string[] = [];
|
||||
const selectedConcepts: string[] = [];
|
||||
|
||||
// Look for selectedTools array
|
||||
const toolsMatch = jsonString.match(/"selectedTools"\s*:\s*\[([\s\S]*?)\]/i);
|
||||
if (toolsMatch) {
|
||||
const toolsContent = toolsMatch[1];
|
||||
const toolMatches = toolsContent.match(/"([^"]+)"/g);
|
||||
if (toolMatches) {
|
||||
selectedTools.push(...toolMatches.map(match => match.replace(/"/g, '')));
|
||||
}
|
||||
}
|
||||
|
||||
// Look for selectedConcepts array
|
||||
const conceptsMatch = jsonString.match(/"selectedConcepts"\s*:\s*\[([\s\S]*?)\]/i);
|
||||
if (conceptsMatch) {
|
||||
const conceptsContent = conceptsMatch[1];
|
||||
const conceptMatches = conceptsContent.match(/"([^"]+)"/g);
|
||||
if (conceptMatches) {
|
||||
selectedConcepts.push(...conceptMatches.map(match => match.replace(/"/g, '')));
|
||||
}
|
||||
}
|
||||
|
||||
// If we couldn't parse arrays separately, fall back to generic name extraction
|
||||
if (selectedTools.length === 0 && selectedConcepts.length === 0) {
|
||||
const allMatches = jsonString.match(/"([^"]+)"/g);
|
||||
if (allMatches) {
|
||||
const possibleNames = allMatches
|
||||
.map(match => match.replace(/"/g, ''))
|
||||
.filter(name =>
|
||||
name.length > 2 &&
|
||||
!['selectedTools', 'selectedConcepts', 'reasoning'].includes(name) &&
|
||||
!name.includes(':') && // Avoid JSON keys
|
||||
!name.match(/^\d+$/) // Avoid pure numbers
|
||||
)
|
||||
.slice(0, 15);
|
||||
|
||||
// Assume all recovered names are tools (since concepts are usually fewer)
|
||||
selectedTools.push(...possibleNames);
|
||||
}
|
||||
}
|
||||
|
||||
if (selectedTools.length > 0 || selectedConcepts.length > 0) {
|
||||
console.log(`[AI PIPELINE] Recovery successful: ${selectedTools.length} tools, ${selectedConcepts.length} concepts`);
|
||||
return {
|
||||
selectedTools,
|
||||
selectedConcepts,
|
||||
reasoning: 'Recovered from malformed JSON response'
|
||||
};
|
||||
}
|
||||
}
|
||||
|
||||
console.error('[AI PIPELINE] All recovery attempts failed');
|
||||
return fallback;
|
||||
}
|
||||
}
|
||||
@ -491,7 +572,29 @@ class ImprovedMicroTaskAIPipeline {
|
||||
) {
|
||||
const selectionStart = Date.now();
|
||||
|
||||
const toolsWithFullData = candidateTools.map((tool: any) => ({
|
||||
const candidateMethods = candidateTools.filter(tool => tool.type === 'method');
|
||||
const candidateSoftware = candidateTools.filter(tool => tool.type === 'software');
|
||||
|
||||
console.log(`[AI PIPELINE] Candidates: ${candidateMethods.length} methods, ${candidateSoftware.length} software, ${candidateConcepts.length} concepts`);
|
||||
|
||||
const methodsWithFullData = candidateMethods.map((tool: any) => ({
|
||||
name: tool.name,
|
||||
type: tool.type,
|
||||
description: tool.description,
|
||||
domains: tool.domains,
|
||||
phases: tool.phases,
|
||||
platforms: tool.platforms || [],
|
||||
tags: tool.tags || [],
|
||||
skillLevel: tool.skillLevel,
|
||||
license: tool.license,
|
||||
accessType: tool.accessType,
|
||||
projectUrl: tool.projectUrl,
|
||||
knowledgebase: tool.knowledgebase,
|
||||
related_concepts: tool.related_concepts || [],
|
||||
related_software: tool.related_software || []
|
||||
}));
|
||||
|
||||
const softwareWithFullData = candidateSoftware.map((tool: any) => ({
|
||||
name: tool.name,
|
||||
type: tool.type,
|
||||
description: tool.description,
|
||||
@ -524,36 +627,72 @@ class ImprovedMicroTaskAIPipeline {
|
||||
let conceptsToSend: any[];
|
||||
|
||||
if (selectionMethod === 'embeddings_candidates') {
|
||||
toolsToSend = toolsWithFullData.slice(0, this.embeddingSelectionLimit);
|
||||
const totalLimit = this.embeddingSelectionLimit;
|
||||
|
||||
const methodLimit = Math.ceil(totalLimit * this.methodSelectionRatio);
|
||||
const softwareLimit = Math.floor(totalLimit * this.softwareSelectionRatio);
|
||||
|
||||
const selectedMethods = methodsWithFullData.slice(0, methodLimit);
|
||||
const selectedSoftware = softwareWithFullData.slice(0, softwareLimit);
|
||||
|
||||
toolsToSend = [...selectedMethods, ...selectedSoftware];
|
||||
|
||||
const remainingCapacity = totalLimit - toolsToSend.length;
|
||||
if (remainingCapacity > 0) {
|
||||
if (methodsWithFullData.length > methodLimit) {
|
||||
toolsToSend.push(...methodsWithFullData.slice(methodLimit, methodLimit + remainingCapacity));
|
||||
} else if (softwareWithFullData.length > softwareLimit) {
|
||||
toolsToSend.push(...softwareWithFullData.slice(softwareLimit, softwareLimit + remainingCapacity));
|
||||
}
|
||||
}
|
||||
|
||||
conceptsToSend = conceptsWithFullData.slice(0, this.embeddingConceptsLimit);
|
||||
|
||||
console.log(`[AI PIPELINE] Embeddings enabled: sending top ${toolsToSend.length} similarity-ordered tools`);
|
||||
console.log(`[AI PIPELINE] Balanced selection: ${selectedMethods.length} methods, ${selectedSoftware.length} software, ${conceptsToSend.length} concepts`);
|
||||
console.log(`[AI PIPELINE] Method names: ${selectedMethods.slice(0, 5).map(m => m.name).join(', ')}${selectedMethods.length > 5 ? '...' : ''}`);
|
||||
console.log(`[AI PIPELINE] Software names: ${selectedSoftware.slice(0, 5).map(s => s.name).join(', ')}${selectedSoftware.length > 5 ? '...' : ''}`);
|
||||
console.log(`[AI PIPELINE] Concept names: ${conceptsToSend.map(c => c.name).join(', ')}`);
|
||||
|
||||
} else {
|
||||
const maxTools = this.noEmbeddingsToolLimit > 0 ?
|
||||
Math.min(this.noEmbeddingsToolLimit, candidateTools.length) :
|
||||
candidateTools.length;
|
||||
const maxTools = this.noEmbeddingsToolLimit > 0 ? this.noEmbeddingsToolLimit : 25;
|
||||
const maxConcepts = this.noEmbeddingsConceptLimit > 0 ? this.noEmbeddingsConceptLimit : 10;
|
||||
|
||||
const maxConcepts = this.noEmbeddingsConceptLimit > 0 ?
|
||||
Math.min(this.noEmbeddingsConceptLimit, candidateConcepts.length) :
|
||||
candidateConcepts.length;
|
||||
const methodLimit = Math.ceil(maxTools * 0.4);
|
||||
const softwareLimit = Math.floor(maxTools * 0.5);
|
||||
|
||||
const selectedMethods = methodsWithFullData.slice(0, methodLimit);
|
||||
const selectedSoftware = softwareWithFullData.slice(0, softwareLimit);
|
||||
|
||||
toolsToSend = [...selectedMethods, ...selectedSoftware];
|
||||
|
||||
const remainingCapacity = maxTools - toolsToSend.length;
|
||||
if (remainingCapacity > 0) {
|
||||
if (methodsWithFullData.length > methodLimit) {
|
||||
toolsToSend.push(...methodsWithFullData.slice(methodLimit, methodLimit + remainingCapacity));
|
||||
} else if (softwareWithFullData.length > softwareLimit) {
|
||||
toolsToSend.push(...softwareWithFullData.slice(softwareLimit, softwareLimit + remainingCapacity));
|
||||
}
|
||||
}
|
||||
|
||||
toolsToSend = toolsWithFullData.slice(0, maxTools);
|
||||
conceptsToSend = conceptsWithFullData.slice(0, maxConcepts);
|
||||
|
||||
console.log(`[AI PIPELINE] Embeddings disabled: sending ${toolsToSend.length}/${candidateTools.length} tools (limit: ${this.noEmbeddingsToolLimit || 'none'})`);
|
||||
console.log(`[AI PIPELINE] Balanced selection (no embeddings): ${selectedMethods.length} methods, ${selectedSoftware.length} software, ${conceptsToSend.length} concepts`);
|
||||
}
|
||||
|
||||
const basePrompt = getPrompt('toolSelection', mode, userQuery, selectionMethod, this.maxSelectedItems);
|
||||
const prompt = `${basePrompt}
|
||||
|
||||
VERFÜGBARE TOOLS (mit vollständigen Daten):
|
||||
${JSON.stringify(toolsToSend, null, 2)}
|
||||
VERFÜGBARE TOOLS (${toolsToSend.length} Items - Methoden und Software):
|
||||
${JSON.stringify(toolsToSend, null, 2)}
|
||||
|
||||
VERFÜGBARE KONZEPTE (mit vollständigen Daten):
|
||||
${JSON.stringify(conceptsToSend, null, 2)}`;
|
||||
VERFÜGBARE KONZEPTE (${conceptsToSend.length} Items - theoretisches Wissen):
|
||||
${JSON.stringify(conceptsToSend, null, 2)}
|
||||
|
||||
WICHTIGER HINWEIS: Wähle sowohl aus TOOLS als auch aus KONZEPTEN aus! Konzepte sind essentiell für methodische Fundierung.`;
|
||||
|
||||
const estimatedTokens = this.estimateTokens(prompt);
|
||||
console.log(`[AI PIPELINE] Method: ${selectionMethod}, Tools: ${toolsToSend.length}, Estimated tokens: ~${estimatedTokens}`);
|
||||
console.log(`[AI PIPELINE] Sending to AI: ${toolsToSend.filter(t => t.type === 'method').length} methods, ${toolsToSend.filter(t => t.type === 'software').length} software, ${conceptsToSend.length} concepts`);
|
||||
console.log(`[AI PIPELINE] Estimated tokens: ~${estimatedTokens}`);
|
||||
|
||||
if (estimatedTokens > 35000) {
|
||||
console.warn(`[AI PIPELINE] WARNING: Prompt tokens (${estimatedTokens}) may exceed model limits`);
|
||||
@ -569,43 +708,61 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
|
||||
|
||||
if (this.auditConfig.enabled) {
|
||||
this.addAuditEntry(context, 'selection', 'ai-tool-selection-failed',
|
||||
{ candidateCount: candidateTools.length, mode, prompt: prompt.slice(0, 200) },
|
||||
{ candidateCount: candidateTools.length, candidateConceptsCount: candidateConcepts.length, mode, prompt: prompt.slice(0, 200) },
|
||||
{ error: 'Invalid JSON structure', response: response.slice(0, 200) },
|
||||
10,
|
||||
selectionStart,
|
||||
{ aiModel: this.config.model, selectionMethod, tokensSent: estimatedTokens, toolsSent: toolsToSend.length }
|
||||
{ aiModel: this.config.model, selectionMethod, tokensSent: estimatedTokens, toolsSent: toolsToSend.length, conceptsSent: conceptsToSend.length }
|
||||
);
|
||||
}
|
||||
|
||||
throw new Error('AI selection failed to return valid tool selection');
|
||||
throw new Error('AI selection failed to return valid tool and concept selection');
|
||||
}
|
||||
|
||||
const totalSelected = result.selectedTools.length + result.selectedConcepts.length;
|
||||
if (totalSelected === 0) {
|
||||
console.error('[AI PIPELINE] AI selection returned no tools');
|
||||
console.error('[AI PIPELINE] AI selection returned no tools or concepts');
|
||||
throw new Error('AI selection returned empty selection');
|
||||
}
|
||||
|
||||
console.log(`[AI PIPELINE] AI selected: ${result.selectedTools.length} tools, ${result.selectedConcepts.length} concepts from ${toolsToSend.length} candidates`);
|
||||
|
||||
const selectedTools = candidateTools.filter(tool => result.selectedTools.includes(tool.name));
|
||||
const selectedConcepts = candidateConcepts.filter(concept => result.selectedConcepts.includes(concept.name));
|
||||
// Create lookup maps for efficient filtering
|
||||
const toolsMap = new Map<string, any>(candidateTools.map(tool => [tool.name, tool]));
|
||||
const conceptsMap = new Map<string, any>(candidateConcepts.map(concept => [concept.name, concept]));
|
||||
|
||||
const selectedTools = result.selectedTools
|
||||
.map(name => toolsMap.get(name))
|
||||
.filter((tool): tool is any => tool !== undefined);
|
||||
|
||||
const selectedConcepts = result.selectedConcepts
|
||||
.map(name => conceptsMap.get(name))
|
||||
.filter((concept): concept is any => concept !== undefined);
|
||||
|
||||
const selectedMethods = selectedTools.filter(t => t.type === 'method');
|
||||
const selectedSoftware = selectedTools.filter(t => t.type === 'software');
|
||||
|
||||
console.log(`[AI PIPELINE] AI selected: ${selectedMethods.length} methods, ${selectedSoftware.length} software, ${selectedConcepts.length} concepts`);
|
||||
console.log(`[AI PIPELINE] Selection balance: ${((selectedMethods.length / (selectedTools.length || 1)) * 100).toFixed(0)}% methods`);
|
||||
console.log(`[AI PIPELINE] Selected tool names: ${selectedTools.map(t => t.name).join(', ')}`);
|
||||
console.log(`[AI PIPELINE] Selected concept names: ${selectedConcepts.map(c => c.name).join(', ')}`);
|
||||
|
||||
if (this.auditConfig.enabled) {
|
||||
const confidence = this.calculateSelectionConfidence(result, candidateTools.length);
|
||||
const confidence = this.calculateSelectionConfidence(result, candidateTools.length + candidateConcepts.length);
|
||||
|
||||
this.addAuditEntry(context, 'selection', 'ai-tool-selection',
|
||||
{ candidateCount: candidateTools.length, mode, promptLength: prompt.length },
|
||||
{ candidateCount: candidateTools.length, candidateConceptsCount: candidateConcepts.length, mode, promptLength: prompt.length },
|
||||
{
|
||||
selectedToolCount: result.selectedTools.length,
|
||||
selectedConceptCount: result.selectedConcepts.length,
|
||||
selectedMethodCount: selectedMethods.length,
|
||||
selectedSoftwareCount: selectedSoftware.length,
|
||||
selectedConceptCount: selectedConcepts.length,
|
||||
reasoning: result.reasoning?.slice(0, 200) + '...',
|
||||
finalToolNames: selectedTools.map(t => t.name),
|
||||
selectionEfficiency: `${toolsToSend.length} → ${result.selectedTools.length}`
|
||||
finalConceptNames: selectedConcepts.map(c => c.name),
|
||||
methodBalance: `${((selectedMethods.length / (selectedTools.length || 1)) * 100).toFixed(0)}%`,
|
||||
conceptsSelected: selectedConcepts.length > 0
|
||||
},
|
||||
confidence,
|
||||
selectionStart,
|
||||
{ aiModel: this.config.model, selectionMethod, promptTokens: estimatedTokens, toolsSent: toolsToSend.length }
|
||||
{ aiModel: this.config.model, selectionMethod, promptTokens: estimatedTokens, toolsSent: toolsToSend.length, conceptsSent: conceptsToSend.length }
|
||||
);
|
||||
}
|
||||
|
||||
@ -619,7 +776,7 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
|
||||
|
||||
if (this.auditConfig.enabled) {
|
||||
this.addAuditEntry(context, 'selection', 'ai-tool-selection-error',
|
||||
{ candidateCount: candidateTools.length, mode },
|
||||
{ candidateCount: candidateTools.length, candidateConceptsCount: candidateConcepts.length, mode },
|
||||
{ error: error.message },
|
||||
5,
|
||||
selectionStart,
|
||||
@ -851,7 +1008,11 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
|
||||
tool.phases && tool.phases.includes(phase.id)
|
||||
);
|
||||
|
||||
console.log(`[AI PIPELINE] Phase ${phase.id} (${phase.name}): Found ${phaseTools.length} matching tools`);
|
||||
console.log(`[AI PIPELINE] Available tools for phase: ${phaseTools.map(t => `${t.name}(${t.type})`).join(', ')}`);
|
||||
|
||||
if (phaseTools.length === 0) {
|
||||
console.log(`[AI PIPELINE] No tools available for phase ${phase.id}, skipping`);
|
||||
return {
|
||||
taskType: 'tool-selection',
|
||||
content: JSON.stringify([]),
|
||||
@ -860,17 +1021,38 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
|
||||
};
|
||||
}
|
||||
|
||||
const phaseMethods = phaseTools.filter(t => t.type === 'method');
|
||||
const phaseSoftware = phaseTools.filter(t => t.type === 'software');
|
||||
|
||||
console.log(`[AI PIPELINE] Phase ${phase.id}: ${phaseMethods.length} methods, ${phaseSoftware.length} software`);
|
||||
|
||||
const prompt = getPrompt('phaseToolSelection', context.userQuery, phase, phaseTools);
|
||||
|
||||
const result = await this.callMicroTaskAI(prompt, context, 1000);
|
||||
|
||||
if (result.success) {
|
||||
console.log(`[AI PIPELINE] Phase ${phase.id} AI response length: ${result.content.length}`);
|
||||
console.log(`[AI PIPELINE] Phase ${phase.id} AI response preview: ${result.content.slice(0, 200)}...`);
|
||||
|
||||
const selections = this.safeParseJSON(result.content, []);
|
||||
|
||||
if (Array.isArray(selections)) {
|
||||
const validSelections = selections.filter((sel: any) =>
|
||||
sel.toolName && phaseTools.some((tool: any) => tool.name === sel.toolName)
|
||||
);
|
||||
console.log(`[AI PIPELINE] Phase ${phase.id}: Parsed ${selections.length} selections from AI`);
|
||||
|
||||
const validSelections = selections.filter((sel: any) => {
|
||||
if (!sel.toolName) return false;
|
||||
|
||||
let matchingTool = phaseTools.find((tool: any) => tool.name === sel.toolName);
|
||||
|
||||
const isValid = !!matchingTool;
|
||||
if (!isValid) {
|
||||
console.warn(`[AI PIPELINE] Phase ${phase.id}: Invalid selection: ${JSON.stringify(sel)}`);
|
||||
console.warn(`[AI PIPELINE] Phase ${phase.id}: Available tool names: ${phaseTools.map(t => t.name).join(', ')}`);
|
||||
}
|
||||
return isValid;
|
||||
});
|
||||
|
||||
console.log(`[AI PIPELINE] Phase ${phase.id}: ${validSelections.length} valid selections after filtering`);
|
||||
|
||||
validSelections.forEach((sel: any) => {
|
||||
const tool = phaseTools.find((t: any) => t.name === sel.toolName);
|
||||
@ -880,25 +1062,35 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
|
||||
|
||||
const priority = this.derivePriorityFromScore(taskRelevance);
|
||||
|
||||
console.log(`[AI PIPELINE] Phase ${phase.id}: Adding ${tool.name} (${tool.type}) with priority ${priority}, relevance ${taskRelevance}%`);
|
||||
|
||||
this.addToolToSelection(context, tool, phase.id, priority, sel.justification, taskRelevance, sel.limitations);
|
||||
}
|
||||
});
|
||||
|
||||
this.addAuditEntry(context, 'micro-task', 'phase-tool-selection',
|
||||
{ phase: phase.id, availableTools: phaseTools.length },
|
||||
{ phase: phase.id, availableTools: phaseTools.length, availableMethods: phaseMethods.length, availableSoftware: phaseSoftware.length },
|
||||
{
|
||||
validSelections: validSelections.length,
|
||||
selectedTools: validSelections.map(s => ({
|
||||
name: s.toolName,
|
||||
taskRelevance: s.taskRelevance,
|
||||
derivedPriority: this.derivePriorityFromScore(s.taskRelevance)
|
||||
}))
|
||||
})),
|
||||
methodsSelected: validSelections.filter(s => {
|
||||
const tool = phaseTools.find(t => t.name === s.toolName);
|
||||
return tool && tool.type === 'method';
|
||||
}).length
|
||||
},
|
||||
validSelections.length > 0 ? 75 : 30,
|
||||
Date.now() - result.processingTimeMs,
|
||||
{ phaseName: phase.name, comparativeEvaluation: true, priorityDerived: true }
|
||||
);
|
||||
} else {
|
||||
console.error(`[AI PIPELINE] Phase ${phase.id}: Failed to parse selections as array:`, selections);
|
||||
}
|
||||
} else {
|
||||
console.error(`[AI PIPELINE] Phase ${phase.id}: AI call failed:`, result.error);
|
||||
}
|
||||
|
||||
return result;
|
||||
@ -1057,6 +1249,32 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
|
||||
return 'low';
|
||||
}
|
||||
|
||||
private async performAISelection(
|
||||
filteredData: any,
|
||||
userQuery: string,
|
||||
mode: string,
|
||||
context: AnalysisContext
|
||||
): Promise<{ tools: any[], concepts: any[] }> {
|
||||
const selectionStart = Date.now();
|
||||
|
||||
// Call the existing aiSelectionWithFullData
|
||||
const result = await this.aiSelectionWithFullData(
|
||||
userQuery,
|
||||
filteredData.tools,
|
||||
filteredData.concepts,
|
||||
mode,
|
||||
embeddingsService.isEnabled() ? 'embeddings_candidates' : 'full_dataset',
|
||||
context
|
||||
);
|
||||
|
||||
console.log(`[AI PIPELINE] AI selection complete: ${result.selectedTools.length} tools, ${result.selectedConcepts.length} concepts`);
|
||||
|
||||
return {
|
||||
tools: result.selectedTools,
|
||||
concepts: result.selectedConcepts
|
||||
};
|
||||
}
|
||||
|
||||
async processQuery(userQuery: string, mode: string): Promise<AnalysisResult> {
|
||||
const startTime = Date.now();
|
||||
let completeTasks = 0;
|
||||
@ -1078,15 +1296,34 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
|
||||
currentContextLength: 0,
|
||||
seenToolNames: new Set<string>(),
|
||||
auditTrail: [],
|
||||
embeddingsSimilarities: new Map<string, number>()
|
||||
embeddingsSimilarities: new Map<string, number>(),
|
||||
// Add this new property to store AI selections
|
||||
aiSelectedTools: [],
|
||||
aiSelectedConcepts: []
|
||||
};
|
||||
|
||||
// Get embedding-filtered candidates
|
||||
const filteredData = await this.getIntelligentCandidates(userQuery, toolsData, mode, context);
|
||||
context.filteredData = filteredData;
|
||||
|
||||
// IMPORTANT: Now do the AI selection from those candidates
|
||||
const aiSelection = await this.performAISelection(filteredData, userQuery, mode, context);
|
||||
|
||||
// Store the AI's selections in context
|
||||
context.aiSelectedTools = aiSelection.tools;
|
||||
context.aiSelectedConcepts = aiSelection.concepts;
|
||||
|
||||
// Update filteredData to only include what AI selected
|
||||
context.filteredData = {
|
||||
tools: aiSelection.tools,
|
||||
concepts: aiSelection.concepts,
|
||||
domains: filteredData.domains,
|
||||
phases: filteredData.phases,
|
||||
'domain-agnostic-software': filteredData['domain-agnostic-software']
|
||||
};
|
||||
|
||||
this.mergeTemporaryAuditEntries(context);
|
||||
|
||||
console.log(`[AI PIPELINE] Starting micro-tasks with ${filteredData.tools.length} tools visible`);
|
||||
console.log(`[AI PIPELINE] Starting micro-tasks with ${context.filteredData.tools.length} AI-selected tools`);
|
||||
|
||||
this.addAuditEntry(context, 'initialization', 'pipeline-start',
|
||||
{ userQuery, mode, toolsDataLoaded: !!toolsData },
|
||||
@ -1110,6 +1347,27 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
|
||||
|
||||
if (mode === 'workflow') {
|
||||
const phases = toolsData.phases || [];
|
||||
|
||||
console.log(`[AI PIPELINE] Debug: Starting phase-specific selection with ${context.filteredData.tools.length} AI-selected tools`);
|
||||
console.log(`[AI PIPELINE] Debug: Available phases: ${phases.map(p => p.id).join(', ')}`);
|
||||
|
||||
context.filteredData.tools.forEach(tool => {
|
||||
console.log(`[AI PIPELINE] Debug: ${tool.name} (${tool.type}) - phases: ${tool.phases?.join(', ') || 'NO PHASES'}`);
|
||||
});
|
||||
|
||||
phases.forEach(phase => {
|
||||
const matchingTools = context.filteredData.tools.filter(tool =>
|
||||
tool.phases && tool.phases.includes(phase.id)
|
||||
);
|
||||
const matchingMethods = matchingTools.filter(t => t.type === 'method');
|
||||
const matchingSoftware = matchingTools.filter(t => t.type === 'software');
|
||||
|
||||
console.log(`[AI PIPELINE] Debug: Phase ${phase.id} has ${matchingTools.length} matching tools (${matchingMethods.length} methods, ${matchingSoftware.length} software)`);
|
||||
if (matchingTools.length > 0) {
|
||||
console.log(`[AI PIPELINE] Debug: Phase ${phase.id} tools: ${matchingTools.map(t => `${t.name}(${t.type})`).join(', ')}`);
|
||||
}
|
||||
});
|
||||
|
||||
for (const phase of phases) {
|
||||
const toolSelectionResult = await this.selectToolsForPhase(context, phase);
|
||||
if (toolSelectionResult.success) completeTasks++; else failedTasks++;
|
||||
@ -1176,6 +1434,24 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
|
||||
private buildRecommendation(context: AnalysisContext, mode: string, finalContent: string): any {
|
||||
const isWorkflow = mode === 'workflow';
|
||||
|
||||
console.log(`[AI PIPELINE] Building recommendation for ${mode} mode`);
|
||||
console.log(`[AI PIPELINE] Selected tools count: ${context.selectedTools?.length || 0}`);
|
||||
|
||||
if (context.selectedTools && context.selectedTools.length > 0) {
|
||||
const methods = context.selectedTools.filter(st => st.tool.type === 'method');
|
||||
const software = context.selectedTools.filter(st => st.tool.type === 'software');
|
||||
|
||||
console.log(`[AI PIPELINE] Final selection breakdown: ${methods.length} methods, ${software.length} software`);
|
||||
console.log(`[AI PIPELINE] Method names: ${methods.map(m => m.tool.name).join(', ')}`);
|
||||
console.log(`[AI PIPELINE] Software names: ${software.map(s => s.tool.name).join(', ')}`);
|
||||
|
||||
context.selectedTools.forEach((st, index) => {
|
||||
console.log(`[AI PIPELINE] Selected tool ${index + 1}: ${st.tool.name} (${st.tool.type}) - Phase: ${st.phase}, Priority: ${st.priority}`);
|
||||
});
|
||||
} else {
|
||||
console.warn(`[AI PIPELINE] WARNING: No tools in selectedTools array!`);
|
||||
}
|
||||
|
||||
const base = {
|
||||
[isWorkflow ? 'scenario_analysis' : 'problem_analysis']:
|
||||
isWorkflow ? context.scenarioAnalysis : context.problemAnalysis,
|
||||
@ -1201,7 +1477,7 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
|
||||
);
|
||||
|
||||
this.addAuditEntry(context, 'validation', 'confidence-scoring',
|
||||
{ toolName: st.tool.name, phase: st.phase },
|
||||
{ toolName: st.tool.name, toolType: st.tool.type, phase: st.phase },
|
||||
{
|
||||
overall: confidence.overall,
|
||||
components: {
|
||||
@ -1216,6 +1492,7 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
|
||||
|
||||
return {
|
||||
name: st.tool.name,
|
||||
type: st.tool.type,
|
||||
phase: st.phase,
|
||||
priority: st.priority,
|
||||
justification: st.justification || `Empfohlen für ${st.phase}`,
|
||||
@ -1225,6 +1502,11 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
|
||||
};
|
||||
}) || [];
|
||||
|
||||
console.log(`[AI PIPELINE] Final workflow recommendations: ${recommendedToolsWithConfidence.length} tools`);
|
||||
const finalMethods = recommendedToolsWithConfidence.filter(r => r.type === 'method');
|
||||
const finalSoftware = recommendedToolsWithConfidence.filter(r => r.type === 'software');
|
||||
console.log(`[AI PIPELINE] Final breakdown: ${finalMethods.length} methods, ${finalSoftware.length} software`);
|
||||
|
||||
return {
|
||||
...base,
|
||||
recommended_tools: recommendedToolsWithConfidence,
|
||||
@ -1241,7 +1523,7 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
|
||||
);
|
||||
|
||||
this.addAuditEntry(context, 'validation', 'confidence-scoring',
|
||||
{ toolName: st.tool.name, rank: st.tool.evaluation?.rank || 1 },
|
||||
{ toolName: st.tool.name, toolType: st.tool.type, rank: st.tool.evaluation?.rank || 1 },
|
||||
{
|
||||
overall: confidence.overall,
|
||||
suitabilityAlignment: st.priority === 'high' && confidence.overall >= this.confidenceConfig.highThreshold,
|
||||
@ -1254,6 +1536,7 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
|
||||
|
||||
return {
|
||||
name: st.tool.name,
|
||||
type: st.tool.type,
|
||||
rank: st.tool.evaluation?.rank || 1,
|
||||
suitability_score: st.priority,
|
||||
detailed_explanation: st.tool.evaluation?.detailed_explanation || '',
|
||||
@ -1267,6 +1550,8 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
|
||||
};
|
||||
}) || [];
|
||||
|
||||
console.log(`[AI PIPELINE] Final tool recommendations: ${recommendedToolsWithConfidence.length} tools`);
|
||||
|
||||
return {
|
||||
...base,
|
||||
recommended_tools: recommendedToolsWithConfidence,
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user