fix tool mode ai pipiline logic
This commit is contained in:
overcuriousity
parent
4ee1cc4984
commit
b14ca1d243
@ -100,17 +100,11 @@ AI_SOFTWARE_SELECTION_RATIO=0.5 # 50% software tools (increase for more tool re
|
||||
|
||||
# AI selection limits
|
||||
AI_MAX_SELECTED_ITEMS=25
|
||||
AI_MAX_TOOLS_TO_ANALYZE=20
|
||||
AI_MAX_CONCEPTS_TO_ANALYZE=10
|
||||
|
||||
# Efficiency thresholds
|
||||
AI_EMBEDDINGS_MIN_TOOLS=8
|
||||
AI_EMBEDDINGS_MAX_REDUCTION_RATIO=0.75
|
||||
|
||||
# Fallback limits when embeddings are disabled
|
||||
AI_NO_EMBEDDINGS_TOOL_LIMIT=25
|
||||
AI_NO_EMBEDDINGS_CONCEPT_LIMIT=10
|
||||
|
||||
# === Rate Limiting & Timing ===
|
||||
AI_MICRO_TASK_TOTAL_LIMIT=30
|
||||
AI_MICRO_TASK_DELAY_MS=500
|
||||
|
||||
@ -3,14 +3,16 @@
|
||||
export const AI_PROMPTS = {
|
||||
|
||||
toolSelection: (mode: string, userQuery: string, maxSelectedItems: number) => {
|
||||
const modeInstruction = mode === 'workflow'
|
||||
? '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.';
|
||||
const modeInstruction =
|
||||
mode === 'workflow'
|
||||
? '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 `Du bist ein DFIR-Experte. Wähle die BESTEN Items aus dem vorgefilterten Set.
|
||||
|
||||
AUSWAHLMETHODE:
|
||||
'✓ Semantisch relevante Items bereits vorgefiltert\n✓ Wähle die BESTEN für die konkrete Aufgabe'}
|
||||
✓ Semantisch relevante Items bereits vorgefiltert
|
||||
✓ Wähle die BESTEN für die konkrete Aufgabe
|
||||
|
||||
${modeInstruction}
|
||||
|
||||
@ -22,28 +24,27 @@ VERFÜGBARE ITEM-TYPEN:
|
||||
|
||||
AUSWAHLSTRATEGIE:
|
||||
1. **ERSTE PRIORITÄT: Relevanz zur Anfrage**
|
||||
- Direkt anwendbar auf das Problem
|
||||
- Löst die Kernherausforderung
|
||||
- Direkt anwendbar auf das Problem
|
||||
- Löst die Kernherausforderung
|
||||
|
||||
2. **ZWEITE PRIORITÄT: Ausgewogene Mischung**
|
||||
- Tools/Methoden für praktische Umsetzung → selectedTools
|
||||
- Konzepte für methodisches Verständnis → selectedConcepts
|
||||
- WICHTIG: Auch Konzepte auswählen, nicht nur Tools!
|
||||
- Tools/Methoden für praktische Umsetzung → selectedTools
|
||||
- Konzepte für methodisches Verständnis → selectedConcepts
|
||||
- WICHTIG: Auch Konzepte auswählen, nicht nur Tools!
|
||||
|
||||
3. **QUALITÄT > QUANTITÄT**
|
||||
- Lieber weniger perfekte Items als viele mittelmäßige
|
||||
- Jedes Item muss begründbar sein
|
||||
- Lieber weniger perfekte Items als viele mittelmäßige
|
||||
- Jedes Item muss begründbar sein
|
||||
|
||||
4. **TASK RELEVANCE REALISM**
|
||||
- Gib realistische Bewertungen (50-85% typisch)
|
||||
- Vermeide übertriebene 90-100% Scores
|
||||
- Nur bei perfekter Übereinstimmung >85%
|
||||
4. **(Skalenharmonisierung – Info)**
|
||||
- Spätere Schritte vergeben "taskRelevance" nach EINHEITLICHER Skala:
|
||||
55–65 (Basis), 66–75 (gut), 76–85 (sehr gut), >85 nur bei perfekter Übereinstimmung.
|
||||
|
||||
AUSWAHLREGELN:
|
||||
- Wähle ${mode === 'workflow' ? '15-25' : '4-10'} Items total, max ${maxSelectedItems}
|
||||
- Wähle ${mode === 'workflow' ? '15–25' : '4–10'} Items total, max ${maxSelectedItems}
|
||||
- BEIDE Arrays füllen: selectedTools UND selectedConcepts
|
||||
- Mindestens 1-2 Konzepte auswählen für methodische Fundierung
|
||||
- Tools: 40% Methoden (type="method"), Rest Software (type="software")
|
||||
- Mindestens 1–2 Konzepte auswählen für methodische Fundierung
|
||||
- Tools: ca. 40% Methoden (type="method"), Rest Software (type="software"), wenn verfügbar
|
||||
|
||||
ANTWORT AUSSCHLIESSLICH IM JSON-FORMAT:
|
||||
{
|
||||
@ -56,26 +57,26 @@ ANTWORT AUSSCHLIESSLICH IM JSON-FORMAT:
|
||||
toolSelectionWithData: (basePrompt: string, toolsToSend: any[], conceptsToSend: any[]) => {
|
||||
return `${basePrompt}
|
||||
|
||||
VERFÜGBARE TOOLS (${toolsToSend.length} Items - Methoden und Software):
|
||||
VERFÜGBARE TOOLS (${toolsToSend.length} Items – Methoden und Software):
|
||||
${JSON.stringify(toolsToSend, null, 2)}
|
||||
|
||||
VERFÜGBARE KONZEPTE (${conceptsToSend.length} Items - theoretisches Wissen):
|
||||
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.
|
||||
|
||||
TASK RELEVANCE GUIDELINES:
|
||||
- 50-65%: Grundlegend relevant, aber nicht optimal
|
||||
- 66-75%: Gut geeignet für die Aufgabe
|
||||
- 76-85%: Sehr gut geeignet, klare Vorteile
|
||||
- 86-100%: NUR für perfekte Übereinstimmung verwenden`;
|
||||
(Hinweis zur späteren Relevanz-Skala – EINHEITLICH ÜBER ALLE MODI)
|
||||
- 55–65: Grundlegend relevant
|
||||
- 66–75: Gut geeignet
|
||||
- 76–85: Sehr gut geeignet
|
||||
- >85 : NUR bei nahezu perfekter Übereinstimmung`;
|
||||
},
|
||||
|
||||
scenarioAnalysis: (isWorkflow: boolean, userQuery: string) => {
|
||||
const analysisType = isWorkflow ? 'Szenario' : 'Problem';
|
||||
const focus = isWorkflow ?
|
||||
'Angriffsvektoren, betroffene Systeme, Zeitkritikalität' :
|
||||
'Kernherausforderung, verfügbare Daten, methodische Anforderungen';
|
||||
const focus = isWorkflow
|
||||
? 'Angriffsvektoren, betroffene Systeme, Zeitkritikalität'
|
||||
: 'Kernherausforderung, verfügbare Daten, methodische Anforderungen';
|
||||
|
||||
return `DFIR-Experte: Analysiere das ${analysisType}.
|
||||
|
||||
@ -88,9 +89,9 @@ Antwort: Fließtext ohne Listen, max 100 Wörter.`;
|
||||
|
||||
investigationApproach: (isWorkflow: boolean, userQuery: string) => {
|
||||
const approachType = isWorkflow ? 'Untersuchungsansatz' : 'Lösungsansatz';
|
||||
const focus = isWorkflow ?
|
||||
'Triage-Prioritäten, Phasenabfolge, Kontaminationsvermeidung' :
|
||||
'Methodenauswahl, Validierung, Integration';
|
||||
const focus = isWorkflow
|
||||
? 'Triage-Prioritäten, Phasenabfolge, Kontaminationsvermeidung'
|
||||
: 'Methodenauswahl, Validierung, Integration';
|
||||
|
||||
return `Entwickle einen ${approachType}.
|
||||
|
||||
@ -102,9 +103,9 @@ Antwort: Fließtext ohne Listen, max 100 Wörter.`;
|
||||
},
|
||||
|
||||
criticalConsiderations: (isWorkflow: boolean, userQuery: string) => {
|
||||
const focus = isWorkflow ?
|
||||
'Beweissicherung vs. Gründlichkeit, Chain of Custody' :
|
||||
'Tool-Validierung, False Positives/Negatives, Qualifikationen';
|
||||
const focus = isWorkflow
|
||||
? 'Beweissicherung vs. Gründlichkeit, Chain of Custody'
|
||||
: 'Tool-Validierung, False Positives/Negatives, Qualifikationen';
|
||||
|
||||
return `Identifiziere kritische Überlegungen.
|
||||
|
||||
@ -118,12 +119,12 @@ Antwort: Fließtext ohne Listen, max 100 Wörter.`;
|
||||
phaseToolSelection: (userQuery: string, phase: any, phaseTools: any[]) => {
|
||||
const methods = phaseTools.filter(t => t.type === 'method');
|
||||
const tools = phaseTools.filter(t => t.type === 'software');
|
||||
|
||||
|
||||
if (phaseTools.length === 0) {
|
||||
return `Keine Methoden/Tools für Phase "${phase.name}" verfügbar. Antworte mit leerem Array: []`;
|
||||
}
|
||||
|
||||
return `Du bist ein DFIR-Experte. Wähle die 2-3 BESTEN Items für Phase "${phase.name}".
|
||||
|
||||
return `Du bist ein DFIR-Experte. Wähle die 2–3 BESTEN Items für Phase "${phase.name}".
|
||||
|
||||
SZENARIO: "${userQuery}"
|
||||
PHASE: ${phase.name} - ${phase.description || ''}
|
||||
@ -131,7 +132,7 @@ PHASE: ${phase.name} - ${phase.description || ''}
|
||||
VERFÜGBARE ITEMS (bereits von KI vorausgewählt):
|
||||
${methods.length > 0 ? `
|
||||
METHODEN (${methods.length}):
|
||||
${methods.map((method: any) =>
|
||||
${methods.map((method: any) =>
|
||||
`- ${method.name}
|
||||
Typ: ${method.type}
|
||||
Beschreibung: ${method.description}
|
||||
@ -142,7 +143,7 @@ ${methods.map((method: any) =>
|
||||
|
||||
${tools.length > 0 ? `
|
||||
SOFTWARE TOOLS (${tools.length}):
|
||||
${tools.map((tool: any) =>
|
||||
${tools.map((tool: any) =>
|
||||
`- ${tool.name}
|
||||
Typ: ${tool.type}
|
||||
Beschreibung: ${tool.description}
|
||||
@ -152,16 +153,16 @@ ${tools.map((tool: any) =>
|
||||
` : 'Keine Software-Tools verfügbar'}
|
||||
|
||||
AUSWAHLREGELN FÜR PHASE "${phase.name}":
|
||||
1. Wähle die 2-3 BESTEN Items für diese spezifische Phase
|
||||
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
|
||||
|
||||
TASK RELEVANCE GUIDELINES:
|
||||
- 60-70%: Grundlegend für diese Phase geeignet
|
||||
- 71-80%: Gut geeignet, klare Phasenrelevanz
|
||||
- 81-90%: Sehr gut geeignet, optimal für Phase
|
||||
- 91-100%: NUR für perfekte Phasenübereinstimmung
|
||||
TASK RELEVANCE – EINHEITLICHE SKALA (GANZZAHL 0–100):
|
||||
- 55–65: Basis/ausreichend für diese Phase
|
||||
- 66–75: Gut geeignet, klare Phasenrelevanz
|
||||
- 76–85: Sehr gut geeignet, optimaler Fit
|
||||
- >85 : Nur bei nahezu perfekter Übereinstimmung
|
||||
|
||||
WICHTIG: Verwende EXAKT die Namen wie oben aufgelistet (ohne Präfixe wie M1./T2.)!
|
||||
|
||||
@ -169,42 +170,50 @@ ANTWORT AUSSCHLIESSLICH IM JSON-FORMAT OHNE JEGLICHEN TEXT AUSSERHALB:
|
||||
[
|
||||
{
|
||||
"toolName": "Exakter Name aus der Liste oben",
|
||||
"taskRelevance": 75,
|
||||
"justification": "Detaillierte Begründung (60-80 Wörter) warum optimal für ${phase.name} - erkläre Anwendung, Vorteile und spezifische Relevanz",
|
||||
"taskRelevance": 0,
|
||||
"justification": "Detaillierte Begründung (60–80 Wörter) warum optimal für ${phase.name} – erkläre Anwendung, Vorteile und spezifische Relevanz",
|
||||
"limitations": ["Mögliche Einschränkung für diese Phase"]
|
||||
}
|
||||
]`;
|
||||
},
|
||||
|
||||
toolEvaluation: (userQuery: string, tool: any, rank: number, taskRelevance: number) => {
|
||||
toolEvaluation: (userQuery: string, tool: any, rank: number) => {
|
||||
const itemType = tool.type === 'method' ? 'Methode' : 'Tool';
|
||||
|
||||
return `Erkläre die Anwendung dieser/dieses ${itemType}.
|
||||
|
||||
return `Erkläre die Anwendung dieser/dieses ${itemType} für die Aufgabe.
|
||||
|
||||
PROBLEM: "${userQuery}"
|
||||
${itemType.toUpperCase()}: ${tool.name} (${taskRelevance}% Eignung)
|
||||
${itemType.toUpperCase()}: ${tool.name}
|
||||
TYP: ${tool.type}
|
||||
|
||||
Bereits als Rang ${rank} bewertet.
|
||||
ANWEISUNGEN:
|
||||
- Beurteile ausschließlich entlang der fachlichen Eignung zum PROBLEM.
|
||||
- Nutze nur die dir vorliegenden Texte/Metadaten (keine externen Annahmen).
|
||||
- Bestimme eine EIGENE ganzzahlige Bewertung "taskRelevance" von 0–100 (ohne %).
|
||||
- Verwende die EINHEITLICHE SKALA:
|
||||
55–65 = Basis/ok · 66–75 = gut · 76–85 = sehr gut · >85 = nur bei nahezu perfekter Übereinstimmung.
|
||||
- Bevorzuge realistische Scores im Bereich 60–80 für gute Fits.
|
||||
|
||||
Gib nur JSON im folgenden Schema zurück. Keine Texte davor oder danach.
|
||||
|
||||
ANTWORT AUSSCHLIESSLICH IM JSON-FORMAT OHNE JEGLICHEN TEXT AUSSERHALB DER JSON-STRUKTUR:
|
||||
{
|
||||
"detailed_explanation": "Warum und wie einsetzen",
|
||||
"implementation_approach": "Konkrete Schritte",
|
||||
"pros": ["Vorteil 1", "Vorteil 2"],
|
||||
"limitations": ["Einschränkung 1"],
|
||||
"alternatives": "Alternative Ansätze"
|
||||
"alternatives": "Alternative Ansätze",
|
||||
"taskRelevance": 0
|
||||
}`;
|
||||
},
|
||||
|
||||
backgroundKnowledgeSelection: (userQuery: string, mode: string, selectedToolNames: string[], availableConcepts: any[]) => {
|
||||
return `Wähle 2-4 relevante Konzepte.
|
||||
return `Wähle 2–4 relevante Konzepte.
|
||||
|
||||
${mode === 'workflow' ? 'SZENARIO' : 'PROBLEM'}: "${userQuery}"
|
||||
AUSGEWÄHLTE TOOLS: ${selectedToolNames.join(', ')}
|
||||
|
||||
VERFÜGBARE KONZEPTE (${availableConcepts.length} KI-kuratiert):
|
||||
${availableConcepts.map((c: any) =>
|
||||
${availableConcepts.map((c: any) =>
|
||||
`- ${c.name}: ${c.description}...`
|
||||
).join('\n')}
|
||||
|
||||
@ -253,7 +262,7 @@ Antwort: Prägnanter Fließtext, knappe Begründung für Nachergänzung. Vermeid
|
||||
candidateTools: any[],
|
||||
candidateConcepts: any[]
|
||||
): string {
|
||||
return `Du bist ein DFIR-Experte. Die initiale KI-Auswahl war zu spezifisch - die Phase "${phase.name}" ist unterrepräsentiert.
|
||||
return `Du bist ein DFIR-Experte. Die initiale KI-Auswahl war zu spezifisch – die Phase "${phase.name}" ist unterrepräsentiert.
|
||||
|
||||
KONTEXT: Die Hauptauswahl hat zu wenige Tools für "${phase.name}" identifiziert. Wähle jetzt ergänzende Tools aus semantischer Nachsuche.
|
||||
|
||||
@ -278,10 +287,10 @@ ${candidateConcepts.map((concept: any) => `
|
||||
` : ''}
|
||||
|
||||
AUSWAHLREGELN FÜR NACHERGÄNZUNG:
|
||||
1. Wähle 1-2 BESTE Methoden/Tools die die ${phase.name}-Phase optimal ergänzen
|
||||
1. Wähle 1–2 BESTE Methoden/Tools die die ${phase.name}-Phase optimal ergänzen
|
||||
2. Methoden/Tools müssen für die ursprüngliche Anfrage relevant sein
|
||||
3. Ergänzen, nicht ersetzen - erweitere die zu spezifische Erstauswahl
|
||||
4. Realistische Task Relevance (70-85% typisch für Nachergänzungen)
|
||||
3. Ergänzen, nicht ersetzen – erweitere die zu spezifische Erstauswahl
|
||||
4. Relevanz-Skala weiterhin EINHEITLICH: 55–65/66–75/76–85/>85 nur bei perfekter Übereinstimmung
|
||||
|
||||
ANTWORT AUSSCHLIESSLICH IM JSON-FORMAT:
|
||||
{
|
||||
@ -292,9 +301,9 @@ ANTWORT AUSSCHLIESSLICH IM JSON-FORMAT:
|
||||
},
|
||||
|
||||
finalRecommendations: (isWorkflow: boolean, userQuery: string, selectedToolNames: string[]) => {
|
||||
const focus = isWorkflow ?
|
||||
'Workflow-Schritte, Best Practices, Objektivität' :
|
||||
'Methodische Überlegungen, Validierung, Qualitätssicherung';
|
||||
const focus = isWorkflow
|
||||
? 'Workflow-Schritte, Best Practices, Objektivität'
|
||||
: 'Methodische Überlegungen, Validierung, Qualitätssicherung';
|
||||
|
||||
return `Erstelle ${isWorkflow ? 'Workflow-Empfehlung' : 'methodische Überlegungen'}.
|
||||
|
||||
@ -313,11 +322,12 @@ export function getPrompt(key: 'scenarioAnalysis', isWorkflow: boolean, userQuer
|
||||
export function getPrompt(key: 'investigationApproach', isWorkflow: boolean, userQuery: string): string;
|
||||
export function getPrompt(key: 'criticalConsiderations', isWorkflow: boolean, userQuery: string): string;
|
||||
export function getPrompt(key: 'phaseToolSelection', userQuery: string, phase: any, phaseTools: any[]): string;
|
||||
export function getPrompt(key: 'toolEvaluation', userQuery: string, tool: any, rank: number, taskRelevance: number): string;
|
||||
export function getPrompt(key: 'toolEvaluation', userQuery: string, tool: any, rank: number): string;
|
||||
export function getPrompt(key: 'backgroundKnowledgeSelection', userQuery: string, mode: string, selectedToolNames: string[], availableConcepts: any[]): string;
|
||||
export function getPrompt(key: 'phaseCompletionReasoning', originalQuery: string, phase: any, selectedToolName: string, tool: any, completionContext: string): string;
|
||||
export function getPrompt(key: 'finalRecommendations', isWorkflow: boolean, userQuery: string, selectedToolNames: string[]): string;
|
||||
export function getPrompt(key: 'generatePhaseCompletionPrompt', originalQuery: string, phase: any, candidateTools: any[], candidateConcepts: any[]): string;
|
||||
|
||||
export function getPrompt(promptKey: keyof typeof AI_PROMPTS, ...args: any[]): string {
|
||||
try {
|
||||
const promptFunction = AI_PROMPTS[promptKey];
|
||||
|
||||
@ -470,15 +470,42 @@ class AIPipeline {
|
||||
pipelineStart: number,
|
||||
toolsDataHash: string
|
||||
): Promise<{ completed: number; failed: number }> {
|
||||
const topTools = context.filteredData.tools.slice(0, 3);
|
||||
|
||||
for (let i = 0; i < topTools.length; i++) {
|
||||
const evaluationResult = await this.evaluateSpecificTool(context, topTools[i], i + 1, pipelineStart, toolsDataHash);
|
||||
// Evaluate ALL candidates handed over by the embeddings pre-filter.
|
||||
const candidates = context.filteredData.tools || [];
|
||||
if (!Array.isArray(candidates) || candidates.length === 0) {
|
||||
return { completed: completedTasks, failed: failedTasks };
|
||||
}
|
||||
|
||||
// Evaluate every candidate (no slicing here)
|
||||
for (let i = 0; i < candidates.length; i++) {
|
||||
const evaluationResult = await this.evaluateSpecificTool(context, candidates[i], i + 1, pipelineStart, toolsDataHash);
|
||||
if (evaluationResult.success) completedTasks++; else failedTasks++;
|
||||
this.trackTokenUsage(evaluationResult.aiUsage);
|
||||
await this.delay(this.config.microTaskDelay);
|
||||
}
|
||||
|
||||
|
||||
// At this point, context.selectedTools may contain 0..N evaluated items (added by evaluateSpecificTool).
|
||||
// Now we sort them by AI-derived taskRelevance (after moderation) and keep ONLY the top 3 for UI.
|
||||
if (Array.isArray(context.selectedTools) && context.selectedTools.length > 0) {
|
||||
context.selectedTools.sort((a: any, b: any) => {
|
||||
const ar = typeof a.taskRelevance === 'number' ? a.taskRelevance : -1;
|
||||
const br = typeof b.taskRelevance === 'number' ? b.taskRelevance : -1;
|
||||
if (br !== ar) return br - ar;
|
||||
|
||||
// tie-breakers without domain heuristics:
|
||||
const aLen = (a.justification || '').length;
|
||||
const bLen = (b.justification || '').length;
|
||||
if (bLen !== aLen) return bLen - aLen;
|
||||
|
||||
const aRank = a.tool?.evaluation?.rank ?? Number.MAX_SAFE_INTEGER;
|
||||
const bRank = b.tool?.evaluation?.rank ?? Number.MAX_SAFE_INTEGER;
|
||||
return aRank - bRank;
|
||||
});
|
||||
|
||||
// Keep top 3 only
|
||||
context.selectedTools = context.selectedTools.slice(0, 3);
|
||||
}
|
||||
|
||||
return { completed: completedTasks, failed: failedTasks };
|
||||
}
|
||||
|
||||
@ -849,68 +876,113 @@ class AIPipeline {
|
||||
toolsDataHash: string
|
||||
): Promise<MicroTaskResult> {
|
||||
const taskStart = Date.now();
|
||||
const existingSelection = context.selectedTools?.find((st: any) => st.tool && st.tool.name === tool.name);
|
||||
const originalTaskRelevance = existingSelection?.taskRelevance || 70;
|
||||
const moderatedTaskRelevance = this.moderateTaskRelevance(originalTaskRelevance);
|
||||
const priority = this.derivePriorityFromScore(moderatedTaskRelevance);
|
||||
|
||||
const prompt = getPrompt('toolEvaluation', context.userQuery, tool, rank, moderatedTaskRelevance);
|
||||
|
||||
// Build prompt WITHOUT any baseline score
|
||||
const prompt = getPrompt('toolEvaluation', context.userQuery, tool, rank);
|
||||
const result = await this.callMicroTaskAI(prompt, context, 'tool-evaluation');
|
||||
|
||||
if (result.success) {
|
||||
const evaluation = JSONParser.safeParseJSON(result.content, {
|
||||
detailed_explanation: 'Evaluation failed',
|
||||
implementation_approach: '',
|
||||
pros: [],
|
||||
limitations: [],
|
||||
alternatives: ''
|
||||
});
|
||||
|
||||
this.addToolToSelection(context, {
|
||||
...tool,
|
||||
evaluation: {
|
||||
...evaluation,
|
||||
rank,
|
||||
task_relevance: moderatedTaskRelevance
|
||||
}
|
||||
}, 'evaluation', priority, evaluation.detailed_explanation, moderatedTaskRelevance, evaluation.limitations);
|
||||
|
||||
const responseConfidence = auditService.calculateAIResponseConfidence(
|
||||
result.content,
|
||||
{ min: 200, max: 800 },
|
||||
'tool-evaluation'
|
||||
);
|
||||
|
||||
const finalConfidence = Math.max(responseConfidence, moderatedTaskRelevance);
|
||||
|
||||
|
||||
if (!result.success) {
|
||||
return result;
|
||||
}
|
||||
|
||||
// Parse strictly; do NOT provide a default with a score.
|
||||
const evaluation = JSONParser.safeParseJSON(result.content, null);
|
||||
|
||||
// Require a numeric score produced by the model; otherwise, don't add this tool.
|
||||
const aiProvided = evaluation && typeof evaluation.taskRelevance === 'number' && Number.isFinite(evaluation.taskRelevance)
|
||||
? Math.round(evaluation.taskRelevance)
|
||||
: null;
|
||||
|
||||
if (aiProvided === null) {
|
||||
// Log the malformed output but avoid injecting a synthetic score.
|
||||
auditService.addAIDecision(
|
||||
'tool-evaluation',
|
||||
prompt,
|
||||
result.content,
|
||||
finalConfidence,
|
||||
`Bewertete Tool "${tool.name}" (Rang ${rank}) - Analysierte Eignung für spezifische Aufgabenstellung mit Fokus auf praktische Anwendbarkeit und methodische Integration`,
|
||||
0,
|
||||
`Bewertung für "${tool.name}" ignoriert: fehlender/ungültiger taskRelevance`,
|
||||
taskStart,
|
||||
{
|
||||
toolsDataHash: toolsDataHash,
|
||||
toolsDataHash,
|
||||
microTaskType: 'tool-evaluation',
|
||||
toolName: tool.name,
|
||||
toolType: tool.type,
|
||||
rank,
|
||||
originalTaskRelevance,
|
||||
moderatedTaskRelevance,
|
||||
responseConfidence,
|
||||
finalConfidence,
|
||||
moderationApplied: originalTaskRelevance !== moderatedTaskRelevance,
|
||||
evaluationParsed: !!evaluation.detailed_explanation,
|
||||
prosCount: evaluation.pros?.length || 0,
|
||||
limitationsCount: evaluation.limitations?.length || 0,
|
||||
evaluationParsed: false,
|
||||
decisionBasis: 'ai-analysis',
|
||||
aiModel: aiService.getConfig().model,
|
||||
...result.aiUsage
|
||||
...(result.aiUsage || {})
|
||||
}
|
||||
);
|
||||
return result;
|
||||
}
|
||||
|
||||
|
||||
const moderatedTaskRelevance = this.moderateTaskRelevance(aiProvided);
|
||||
const priority = this.derivePriorityFromScore(moderatedTaskRelevance);
|
||||
|
||||
// Keep original fields if present; coerce to strings/arrays safely.
|
||||
const detailed_explanation = String(evaluation?.detailed_explanation || '').trim();
|
||||
const implementation_approach = String(evaluation?.implementation_approach || '').trim();
|
||||
const pros = Array.isArray(evaluation?.pros) ? evaluation.pros : [];
|
||||
const limitations = Array.isArray(evaluation?.limitations) ? evaluation.limitations : [];
|
||||
const alternatives = String(evaluation?.alternatives || '').trim();
|
||||
|
||||
this.addToolToSelection(
|
||||
context,
|
||||
{
|
||||
...tool,
|
||||
evaluation: {
|
||||
detailed_explanation,
|
||||
implementation_approach,
|
||||
pros,
|
||||
limitations,
|
||||
alternatives,
|
||||
rank,
|
||||
task_relevance: moderatedTaskRelevance
|
||||
}
|
||||
},
|
||||
'evaluation',
|
||||
priority,
|
||||
detailed_explanation,
|
||||
moderatedTaskRelevance,
|
||||
limitations
|
||||
);
|
||||
|
||||
const responseConfidence = auditService.calculateAIResponseConfidence(
|
||||
result.content,
|
||||
{ min: 200, max: 800 },
|
||||
'tool-evaluation'
|
||||
);
|
||||
|
||||
const finalConfidence = Math.max(responseConfidence, moderatedTaskRelevance);
|
||||
|
||||
auditService.addAIDecision(
|
||||
'tool-evaluation',
|
||||
prompt,
|
||||
result.content,
|
||||
finalConfidence,
|
||||
`Bewertete Tool "${tool.name}" (Rang ${rank}) – AI-Score ${aiProvided}, moderiert ${moderatedTaskRelevance}`,
|
||||
taskStart,
|
||||
{
|
||||
toolsDataHash,
|
||||
microTaskType: 'tool-evaluation',
|
||||
toolName: tool.name,
|
||||
toolType: tool.type,
|
||||
rank,
|
||||
aiProvidedTaskRelevance: aiProvided,
|
||||
moderatedTaskRelevance,
|
||||
responseConfidence,
|
||||
finalConfidence,
|
||||
moderationApplied: aiProvided !== moderatedTaskRelevance,
|
||||
evaluationParsed: true,
|
||||
prosCount: pros.length,
|
||||
limitationsCount: limitations.length,
|
||||
decisionBasis: 'ai-analysis',
|
||||
aiModel: aiService.getConfig().model,
|
||||
...(result.aiUsage || {})
|
||||
}
|
||||
);
|
||||
|
||||
return result;
|
||||
}
|
||||
|
||||
|
||||
@ -12,14 +12,13 @@ export interface ToolSelectionConfig {
|
||||
similarityThreshold: number;
|
||||
embeddingSelectionLimit: number;
|
||||
embeddingConceptsLimit: number;
|
||||
noEmbeddingsToolLimit: number;
|
||||
noEmbeddingsConceptLimit: number;
|
||||
embeddingsMinTools: number;
|
||||
embeddingsMaxReductionRatio: number;
|
||||
methodSelectionRatio: number;
|
||||
softwareSelectionRatio: number;
|
||||
}
|
||||
|
||||
|
||||
export interface SelectionContext {
|
||||
userQuery: string;
|
||||
mode: string;
|
||||
@ -51,14 +50,11 @@ class ToolSelector {
|
||||
similarityThreshold: this.getEnvFloat('AI_SIMILARITY_THRESHOLD', 0.3),
|
||||
embeddingSelectionLimit: this.getEnvInt('AI_EMBEDDING_SELECTION_LIMIT', 30),
|
||||
embeddingConceptsLimit: this.getEnvInt('AI_EMBEDDING_CONCEPTS_LIMIT', 15),
|
||||
noEmbeddingsToolLimit: this.getEnvInt('AI_NO_EMBEDDINGS_TOOL_LIMIT', 25),
|
||||
noEmbeddingsConceptLimit: this.getEnvInt('AI_NO_EMBEDDINGS_CONCEPT_LIMIT', 10),
|
||||
embeddingsMinTools: this.getEnvInt('AI_EMBEDDINGS_MIN_TOOLS', 8),
|
||||
embeddingsMaxReductionRatio: this.getEnvFloat('AI_EMBEDDINGS_MAX_REDUCTION_RATIO', 0.75),
|
||||
methodSelectionRatio: this.getEnvFloat('AI_METHOD_SELECTION_RATIO', 0.4),
|
||||
softwareSelectionRatio: this.getEnvFloat('AI_SOFTWARE_SELECTION_RATIO', 0.5)
|
||||
softwareSelectionRatio: this.getEnvFloat('AI_SOFTWARE_SELECTION_RATIO', 0.5),
|
||||
};
|
||||
|
||||
console.log('[TOOL-SELECTOR] Initialized with config:', this.config);
|
||||
}
|
||||
|
||||
@ -185,43 +181,73 @@ class ToolSelector {
|
||||
): Promise<ToolSelectionResult> {
|
||||
console.log('[TOOL-SELECTOR] Performing AI selection');
|
||||
|
||||
const candidateMethods = candidateTools.filter((tool: any) => tool && tool.type === 'method');
|
||||
const candidateSoftware = candidateTools.filter((tool: any) => tool && tool.type === 'software');
|
||||
const candidateMethods = candidateTools.filter((t: any) => t && t.type === 'method');
|
||||
const candidateSoftware = candidateTools.filter((t: any) => t && t.type === 'software');
|
||||
|
||||
console.log('[TOOL-SELECTOR] Candidates:', candidateMethods.length, 'methods,', candidateSoftware.length, 'software,', candidateConcepts.length, 'concepts');
|
||||
console.log('[TOOL-SELECTOR] Candidates:',
|
||||
candidateMethods.length, 'methods,',
|
||||
candidateSoftware.length, 'software,',
|
||||
candidateConcepts.length, 'concepts'
|
||||
);
|
||||
|
||||
const methodsWithFullData = candidateMethods.map(this.createToolData);
|
||||
const softwareWithFullData = candidateSoftware.map(this.createToolData);
|
||||
const conceptsWithFullData = candidateConcepts.map(this.createConceptData);
|
||||
|
||||
const maxTools = Math.min(this.config.embeddingSelectionLimit, this.config.noEmbeddingsToolLimit);
|
||||
const maxConcepts = Math.min(this.config.embeddingConceptsLimit, this.config.noEmbeddingsConceptLimit);
|
||||
const methodLimit = Math.ceil(maxTools * this.config.methodSelectionRatio);
|
||||
const softwareLimit = Math.floor(maxTools * this.config.softwareSelectionRatio);
|
||||
// Embeddings are always ON → only use embedding limits
|
||||
const maxTools = Math.min(this.config.embeddingSelectionLimit, candidateTools.length);
|
||||
const maxConcepts = Math.min(this.config.embeddingConceptsLimit, candidateConcepts.length);
|
||||
|
||||
const toolsToSend: any[] = [
|
||||
...methodsWithFullData.slice(0, methodLimit),
|
||||
...softwareWithFullData.slice(0, softwareLimit),
|
||||
];
|
||||
// Respect ratios first, then fill the remaining capacity
|
||||
const methodRatio = Math.max(0, Math.min(1, this.config.methodSelectionRatio));
|
||||
const softwareRatio = Math.max(0, Math.min(1, this.config.softwareSelectionRatio));
|
||||
|
||||
const remainingCapacity = maxTools - toolsToSend.length;
|
||||
if (remainingCapacity > 0) {
|
||||
const extraMethods = methodsWithFullData.slice(methodLimit, methodLimit + remainingCapacity);
|
||||
const extraSoftware = softwareWithFullData.slice(softwareLimit, softwareLimit + (remainingCapacity - extraMethods.length));
|
||||
toolsToSend.push(...extraMethods, ...extraSoftware);
|
||||
let methodLimit = Math.round(maxTools * methodRatio);
|
||||
let softwareLimit = Math.round(maxTools * softwareRatio);
|
||||
|
||||
// If rounded sum exceeds maxTools, scale down proportionally
|
||||
if (methodLimit + softwareLimit > maxTools) {
|
||||
const scale = maxTools / (methodLimit + softwareLimit);
|
||||
methodLimit = Math.floor(methodLimit * scale);
|
||||
softwareLimit = Math.floor(softwareLimit * scale);
|
||||
}
|
||||
|
||||
const methodsPrimary = methodsWithFullData.slice(0, methodLimit);
|
||||
const softwarePrimary = softwareWithFullData.slice(0, softwareLimit);
|
||||
|
||||
const toolsToSend: any[] = [...methodsPrimary, ...softwarePrimary];
|
||||
|
||||
// Fill any remaining capacity from whichever pool still has candidates
|
||||
let mIdx = methodsPrimary.length;
|
||||
let sIdx = softwarePrimary.length;
|
||||
|
||||
while (toolsToSend.length < maxTools && (mIdx < methodsWithFullData.length || sIdx < softwareWithFullData.length)) {
|
||||
const remM = methodsWithFullData.length - mIdx;
|
||||
const remS = softwareWithFullData.length - sIdx;
|
||||
|
||||
if (remS >= remM && sIdx < softwareWithFullData.length) {
|
||||
toolsToSend.push(softwareWithFullData[sIdx++]);
|
||||
} else if (mIdx < methodsWithFullData.length) {
|
||||
toolsToSend.push(methodsWithFullData[mIdx++]);
|
||||
} else if (sIdx < softwareWithFullData.length) {
|
||||
toolsToSend.push(softwareWithFullData[sIdx++]);
|
||||
} else {
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
const conceptsToSend = conceptsWithFullData.slice(0, maxConcepts);
|
||||
|
||||
const basePrompt = getPrompt('toolSelection', mode, userQuery, this.config.maxSelectedItems);
|
||||
const prompt = getPrompt('toolSelectionWithData', basePrompt, toolsToSend, conceptsToSend);
|
||||
|
||||
console.log('[TOOL-SELECTOR-DEBUG] maxTools:', maxTools, 'maxConcepts:', maxConcepts);
|
||||
console.log('[TOOL-SELECTOR] Sending to AI:',
|
||||
toolsToSend.filter((t: any) => t.type === 'method').length, 'methods,',
|
||||
toolsToSend.filter((t: any) => t.type === 'software').length, 'software,',
|
||||
conceptsToSend.length, 'concepts'
|
||||
);
|
||||
|
||||
const basePrompt = getPrompt('toolSelection', mode, userQuery, this.config.maxSelectedItems);
|
||||
const prompt = getPrompt('toolSelectionWithData', basePrompt, toolsToSend, conceptsToSend);
|
||||
|
||||
try {
|
||||
const response = await aiService.callAI(prompt);
|
||||
const result = JSONParser.safeParseJSON(response.content, null);
|
||||
@ -250,7 +276,11 @@ class ToolSelector {
|
||||
const selectedMethods = selectedTools.filter((t: any) => t && t.type === 'method');
|
||||
const selectedSoftware = selectedTools.filter((t: any) => t && t.type === 'software');
|
||||
|
||||
console.log('[TOOL-SELECTOR] AI selected:', selectedMethods.length, 'methods,', selectedSoftware.length, 'software,', selectedConcepts.length, 'concepts');
|
||||
console.log('[TOOL-SELECTOR] AI selected:',
|
||||
selectedMethods.length, 'methods,',
|
||||
selectedSoftware.length, 'software,',
|
||||
selectedConcepts.length, 'concepts'
|
||||
);
|
||||
|
||||
const confidence = confidenceScoring.calculateSelectionConfidence(
|
||||
result,
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user