cleanup, prompt centralization

This commit is contained in:
overcuriousity 2025-08-29 14:50:11 +02:00
parent b17458d153
commit dc9f52fb7c
6 changed files with 120 additions and 173 deletions

View File

@ -1130,16 +1130,12 @@ class AIQueryInterface {
const lowConfidenceSteps = auditTrail.filter(entry => (entry.confidence || 0) < 60).length; const lowConfidenceSteps = auditTrail.filter(entry => (entry.confidence || 0) < 60).length;
const mediumConfidenceSteps = auditTrail.length - highConfidenceSteps - lowConfidenceSteps; const mediumConfidenceSteps = auditTrail.length - highConfidenceSteps - lowConfidenceSteps;
// FIX 1: Count actual AI decision actions only
const aiDecisionCount = auditTrail.filter(entry => entry.action === 'ai-decision').length; const aiDecisionCount = auditTrail.filter(entry => entry.action === 'ai-decision').length;
// FIX 2: Count actual similarity search actions, not metadata flags
const embeddingsUsageCount = auditTrail.filter(entry => entry.action === 'similarity-search').length; const embeddingsUsageCount = auditTrail.filter(entry => entry.action === 'similarity-search').length;
// FIX 3: Maintain tool selection count (this was correct)
const toolSelectionCount = auditTrail.filter(entry => entry.action === 'selection-decision').length; const toolSelectionCount = auditTrail.filter(entry => entry.action === 'selection-decision').length;
// Additional diagnostic counts for debugging
const microTaskCount = auditTrail.filter(entry => const microTaskCount = auditTrail.filter(entry =>
entry.action === 'ai-decision' && entry.metadata?.microTaskType entry.action === 'ai-decision' && entry.metadata?.microTaskType
).length; ).length;
@ -1152,7 +1148,6 @@ class AIQueryInterface {
entry.action === 'phase-enhancement' entry.action === 'phase-enhancement'
).length; ).length;
// Enhanced insights with diagnostic information
const keyInsights = []; const keyInsights = [];
const potentialIssues = []; const potentialIssues = [];
@ -1172,7 +1167,6 @@ class AIQueryInterface {
keyInsights.push(`${microTaskCount} spezialisierte Micro-Task-Analysen durchgeführt`); keyInsights.push(`${microTaskCount} spezialisierte Micro-Task-Analysen durchgeführt`);
} }
// Detect mode-specific patterns for validation
if (phaseToolSelectionCount > 0 || phaseEnhancementCount > 0) { if (phaseToolSelectionCount > 0 || phaseEnhancementCount > 0) {
keyInsights.push('Workflow-Modus: Phasenspezifische Analyse durchgeführt'); keyInsights.push('Workflow-Modus: Phasenspezifische Analyse durchgeführt');
} else if (microTaskCount >= 3) { } else if (microTaskCount >= 3) {
@ -1216,10 +1210,9 @@ class AIQueryInterface {
keyInsights.push('Mehrheit der Analyseschritte mit hoher Sicherheit'); keyInsights.push('Mehrheit der Analyseschritte mit hoher Sicherheit');
} }
// Validate expected counts based on mode detection
const isWorkflowMode = phaseToolSelectionCount > 0 || phaseEnhancementCount > 0; const isWorkflowMode = phaseToolSelectionCount > 0 || phaseEnhancementCount > 0;
const expectedMinAI = isWorkflowMode ? 11 : 8; // Workflow: 5 common + 6 phase selections, Tool: 5 common + 3 evaluations const expectedMinAI = isWorkflowMode ? 11 : 8;
const expectedMinEmbeddings = 1; // Both modes should have initial search const expectedMinEmbeddings = 1;
if (aiDecisionCount < expectedMinAI) { if (aiDecisionCount < expectedMinAI) {
potentialIssues.push(`${expectedMinAI - aiDecisionCount} fehlende KI-Entscheidungen für ${isWorkflowMode ? 'Workflow' : 'Tool'}-Modus`); potentialIssues.push(`${expectedMinAI - aiDecisionCount} fehlende KI-Entscheidungen für ${isWorkflowMode ? 'Workflow' : 'Tool'}-Modus`);
@ -1250,7 +1243,6 @@ class AIQueryInterface {
analysisQuality, analysisQuality,
keyInsights, keyInsights,
potentialIssues, potentialIssues,
// Debug information
debugCounts: { debugCounts: {
microTaskCount, microTaskCount,
phaseToolSelectionCount, phaseToolSelectionCount,

View File

@ -16,9 +16,41 @@ STRICTNESS:
`.trim(); `.trim();
export const AI_PROMPTS = { export const AI_PROMPTS = {
// --------------------------------------------------------------------------- enhancementQuestions: (input: string) => {
// Tool/Concept selection (AI pre-pick from embedding-curated set) return `Sie sind DFIR-Experte. Ein Nutzer beschreibt unten ein Szenario/Problem.
// ---------------------------------------------------------------------------
ZIEL:
- Stellen Sie NUR dann 13 präzise Rückfragen, wenn entscheidende forensische Lücken die weitere Analyse/Toolauswahl PHASENREIHENFOLGE oder EVIDENCE-STRATEGIE wesentlich beeinflussen würden.
- Wenn ausreichend abgedeckt: Geben Sie eine leere Liste [] zurück.
PRIORITÄT DER THEMEN (in dieser Reihenfolge prüfen):
1) Available Evidence & Artefakte (z.B. RAM-Dump, Disk-Image, Logs, PCAP, Registry, Cloud/Audit-Logs)
2) Scope/Systems (konkrete Plattformen/Assets/Identitäten/Netzsegmente)
3) Investigation Objectives (Ziele: IOC-Extraktion, Timeline, Impact, Attribution)
4) Timeline/Timeframe (kritische Zeitfenster, Erhalt flüchtiger Daten)
5) Legal & Compliance (Chain of Custody, Aufbewahrung, DSGVO/Branchenvorgaben)
6) Technical Constraints (Ressourcen, Zugriffsrechte, Tooling/EDR)
FRAGEN-QUALITÄT:
- Forensisch spezifisch und entscheidungsrelevant (keine Allgemeinplätze).
- Eine Frage pro Thema, keine Dopplungen.
- Antwortbar vom Nutzer (keine Spekulation, keine Beweise senden-Aufforderungen).
- Maximal 18 Wörter, endet mit "?".
VALIDIERUNG:
- Stellen Sie NUR Fragen zu Themen, die im Nutzertext NICHT hinreichend konkret beantwortet sind (keine Wiederholung bereits gegebener Details).
- Wenn alle priorisierten Themen ausreichend sind [].
ANTWORTFORMAT (NUR JSON, KEIN ZUSÄTZLICHER TEXT):
[
"präzise Frage 1?",
"präzise Frage 2?",
"präzise Frage 3?"
]
NUTZER-EINGABE:
${input}`.trim();
},
toolSelection: (mode: string, userQuery: string, maxSelectedItems: number) => { toolSelection: (mode: string, userQuery: string, maxSelectedItems: number) => {
const modeInstruction = const modeInstruction =
mode === 'workflow' mode === 'workflow'
@ -80,9 +112,6 @@ ${RELEVANCE_RUBRIC}
${STRICTNESS}`; ${STRICTNESS}`;
}, },
// ---------------------------------------------------------------------------
// Contextual analyses (keine JSON-Ausgabe nötig; kurzer Fließtext)
// ---------------------------------------------------------------------------
scenarioAnalysis: (isWorkflow: boolean, userQuery: string) => { scenarioAnalysis: (isWorkflow: boolean, userQuery: string) => {
const analysisType = isWorkflow ? 'Szenario' : 'Problem'; const analysisType = isWorkflow ? 'Szenario' : 'Problem';
const focus = isWorkflow const focus = isWorkflow
@ -124,9 +153,6 @@ Fokus: ${focus}
Antwort: Fließtext, max 100 Wörter.`; Antwort: Fließtext, max 100 Wörter.`;
}, },
// ---------------------------------------------------------------------------
// Phase-specific selection (workflow mode substep)
// ---------------------------------------------------------------------------
phaseToolSelection: (userQuery: string, phase: any, phaseTools: any[]) => { phaseToolSelection: (userQuery: string, phase: any, phaseTools: any[]) => {
const methods = phaseTools.filter(t => t.type === 'method'); const methods = phaseTools.filter(t => t.type === 'method');
const tools = phaseTools.filter(t => t.type === 'software'); const tools = phaseTools.filter(t => t.type === 'software');
@ -183,9 +209,6 @@ ANTWORT (NUR JSON):
]`; ]`;
}, },
// ---------------------------------------------------------------------------
// Per-item evaluation (used in tool mode & elsewhere)
// ---------------------------------------------------------------------------
toolEvaluation: (userQuery: string, tool: any, rank: number) => { toolEvaluation: (userQuery: string, tool: any, rank: number) => {
const itemType = tool.type === 'method' ? 'Methode' : 'Tool'; const itemType = tool.type === 'method' ? 'Methode' : 'Tool';
@ -215,9 +238,6 @@ ANTWORT (NUR JSON):
}`; }`;
}, },
// ---------------------------------------------------------------------------
// Background knowledge (concepts)
// ---------------------------------------------------------------------------
backgroundKnowledgeSelection: (userQuery: string, mode: string, selectedToolNames: string[], availableConcepts: any[]) => { backgroundKnowledgeSelection: (userQuery: string, mode: string, selectedToolNames: string[], availableConcepts: any[]) => {
return `Wähle 24 Konzepte, die das Verständnis/den Einsatz der ausgewählten Tools verbessern. return `Wähle 24 Konzepte, die das Verständnis/den Einsatz der ausgewählten Tools verbessern.
@ -242,9 +262,6 @@ ANTWORT (NUR JSON):
]`; ]`;
}, },
// ---------------------------------------------------------------------------
// Phase completion (underrepresented phase fix)
// ---------------------------------------------------------------------------
phaseCompletionReasoning: ( phaseCompletionReasoning: (
originalQuery: string, originalQuery: string,
phase: any, phase: any,
@ -306,9 +323,6 @@ ANTWORT (NUR JSON):
}`; }`;
}, },
// ---------------------------------------------------------------------------
// Final synthesis (short prose, no JSON needed)
// ---------------------------------------------------------------------------
finalRecommendations: (isWorkflow: boolean, userQuery: string, selectedToolNames: string[]) => { finalRecommendations: (isWorkflow: boolean, userQuery: string, selectedToolNames: string[]) => {
const focus = isWorkflow const focus = isWorkflow
? 'Knappe Workflow-Schritte & Best Practices; neutral formulieren' ? 'Knappe Workflow-Schritte & Best Practices; neutral formulieren'
@ -324,7 +338,7 @@ Antwort: Fließtext, max ${isWorkflow ? '100' : '80'} Wörter. Keine Liste.`;
} }
} as const; } as const;
// Overloads export function getPrompt(key: 'enhancementQuestions', input: string): string;
export function getPrompt(key: 'toolSelection', mode: string, userQuery: string, maxSelectedItems: number): string; export function getPrompt(key: 'toolSelection', mode: string, userQuery: string, maxSelectedItems: number): string;
export function getPrompt(key: 'toolSelectionWithData', basePrompt: string, toolsToSend: any[], conceptsToSend: any[]): string; export function getPrompt(key: 'toolSelectionWithData', basePrompt: string, toolsToSend: any[], conceptsToSend: any[]): string;
export function getPrompt(key: 'scenarioAnalysis', isWorkflow: boolean, userQuery: string): string; export function getPrompt(key: 'scenarioAnalysis', isWorkflow: boolean, userQuery: string): string;
@ -337,7 +351,6 @@ export function getPrompt(key: 'phaseCompletionReasoning', originalQuery: string
export function getPrompt(key: 'finalRecommendations', isWorkflow: boolean, userQuery: string, selectedToolNames: 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(key: 'generatePhaseCompletionPrompt', originalQuery: string, phase: any, candidateTools: any[], candidateConcepts: any[]): string;
// Dispatcher
export function getPrompt(promptKey: keyof typeof AI_PROMPTS, ...args: any[]): string { export function getPrompt(promptKey: keyof typeof AI_PROMPTS, ...args: any[]): string {
try { try {
const f = AI_PROMPTS[promptKey]; const f = AI_PROMPTS[promptKey];

View File

@ -273,15 +273,13 @@ import BaseLayout from '../layouts/BaseLayout.astro';
</BaseLayout> </BaseLayout>
<script> <script>
// Load simple-boost from local node_modules instead of CDN // TODO: cleanup
import('simple-boost').then(() => { import('simple-boost').then(() => {
console.log('Simple-boost loaded successfully from local dependencies'); console.log('Simple-boost loaded successfully from local dependencies');
// Setup dynamic amount updating
setupDynamicAmounts(); setupDynamicAmounts();
}).catch(error => { }).catch(error => {
console.error('Failed to load simple-boost:', error); console.error('Failed to load simple-boost:', error);
// Fallback: try to load from node_modules path
const script = document.createElement('script'); const script = document.createElement('script');
script.type = 'module'; script.type = 'module';
script.src = '/node_modules/simple-boost/dist/simple-boost.js'; script.src = '/node_modules/simple-boost/dist/simple-boost.js';
@ -294,7 +292,6 @@ import BaseLayout from '../layouts/BaseLayout.astro';
}); });
function setupDynamicAmounts() { function setupDynamicAmounts() {
// EUR boost button
const eurBoost = document.getElementById('eur-boost'); const eurBoost = document.getElementById('eur-boost');
const eurInput = document.getElementById('eur-amount') as HTMLInputElement; const eurInput = document.getElementById('eur-amount') as HTMLInputElement;
@ -305,7 +302,6 @@ import BaseLayout from '../layouts/BaseLayout.astro';
console.log('EUR amount set to:', amount); console.log('EUR amount set to:', amount);
}); });
// Update on input change for better UX
eurInput.addEventListener('input', () => { eurInput.addEventListener('input', () => {
const amount = parseFloat(eurInput.value) || 0.5; const amount = parseFloat(eurInput.value) || 0.5;
eurBoost.setAttribute('amount', amount.toString()); eurBoost.setAttribute('amount', amount.toString());
@ -315,7 +311,6 @@ import BaseLayout from '../layouts/BaseLayout.astro';
</script> </script>
<style> <style>
/* Custom styling for simple-boost buttons to match your theme */
simple-boost { simple-boost {
--simple-boost-primary: var(--color-warning); --simple-boost-primary: var(--color-warning);
--simple-boost-primary-hover: var(--color-accent); --simple-boost-primary-hover: var(--color-accent);

View File

@ -5,13 +5,53 @@ import { apiError, apiServerError, createAuthErrorResponse } from '../../../util
import { enqueueApiCall } from '../../../utils/rateLimitedQueue.js'; import { enqueueApiCall } from '../../../utils/rateLimitedQueue.js';
import { aiService } from '../../../utils/aiService.js'; import { aiService } from '../../../utils/aiService.js';
import { JSONParser } from '../../../utils/jsonUtils.js'; import { JSONParser } from '../../../utils/jsonUtils.js';
import { getPrompt } from '../../../config/prompts.js';
export const prerender = false; export const prerender = false;
const rateLimitStore = new Map<string, { count: number; resetTime: number }>(); const RATE_LIMIT_WINDOW_MS =
const RATE_LIMIT_WINDOW = 60 * 1000; Number.isFinite(parseInt(process.env.RATE_LIMIT_WINDOW_MS ?? '', 10))
const RATE_LIMIT_MAX = 5; ? parseInt(process.env.RATE_LIMIT_WINDOW_MS!, 10)
: 60_000;
const RATE_LIMIT_MAX =
Number.isFinite(parseInt(process.env.AI_RATE_LIMIT_MAX_REQUESTS ?? '', 10))
? parseInt(process.env.AI_RATE_LIMIT_MAX_REQUESTS!, 10)
: 5;
const INPUT_MIN_CHARS = 40;
const INPUT_MAX_CHARS = 1000;
const Q_MIN_LEN = 15;
const Q_MAX_LEN = 160;
const Q_MAX_COUNT = 3;
const AI_TEMPERATURE = 0.3;
const CLEANER_TEMPERATURE = 0.0;
const rateLimitStore = new Map<string, { count: number; resetTime: number }>();
function checkRateLimit(userId: string): boolean {
const now = Date.now();
const entry = rateLimitStore.get(userId);
if (!entry || now > entry.resetTime) {
rateLimitStore.set(userId, { count: 1, resetTime: now + RATE_LIMIT_WINDOW_MS });
return true;
}
if (entry.count >= RATE_LIMIT_MAX) return false;
entry.count++;
return true;
}
function cleanupExpiredRateLimits(): void {
const now = Date.now();
for (const [userId, entry] of rateLimitStore.entries()) {
if (now > entry.resetTime) rateLimitStore.delete(userId);
}
}
setInterval(cleanupExpiredRateLimits, 5 * 60 * 1000);
/**
* Helpers
*/
function sanitizeInput(input: string): string { function sanitizeInput(input: string): string {
return input return input
.replace(/```[\s\S]*?```/g, '[CODE_BLOCK_REMOVED]') .replace(/```[\s\S]*?```/g, '[CODE_BLOCK_REMOVED]')
@ -19,79 +59,24 @@ function sanitizeInput(input: string): string {
.replace(/\b(system|assistant|user)\s*[:]/gi, '[ROLE_REMOVED]') .replace(/\b(system|assistant|user)\s*[:]/gi, '[ROLE_REMOVED]')
.replace(/\b(ignore|forget|disregard)\s+(previous|all|your)\s+(instructions?|context|rules?)/gi, '[INSTRUCTION_REMOVED]') .replace(/\b(ignore|forget|disregard)\s+(previous|all|your)\s+(instructions?|context|rules?)/gi, '[INSTRUCTION_REMOVED]')
.trim() .trim()
.slice(0, 1000); .slice(0, INPUT_MAX_CHARS);
} }
function checkRateLimit(userId: string): boolean { function stripJsonFences(s: string): string {
const now = Date.now(); return s.replace(/^```json\s*/i, '')
const userLimit = rateLimitStore.get(userId); .replace(/^```\s*/i, '')
.replace(/\s*```\s*$/, '')
if (!userLimit || now > userLimit.resetTime) { .trim();
rateLimitStore.set(userId, { count: 1, resetTime: now + RATE_LIMIT_WINDOW });
return true;
}
if (userLimit.count >= RATE_LIMIT_MAX) {
return false;
}
userLimit.count++;
return true;
}
function cleanupExpiredRateLimits(): void {
const now = Date.now();
for (const [userId, limit] of rateLimitStore.entries()) {
if (now > limit.resetTime) {
rateLimitStore.delete(userId);
}
}
}
setInterval(cleanupExpiredRateLimits, 5 * 60 * 1000);
function createEnhancementPrompt(input: string): string {
return `Sie sind ein DFIR-Experte mit Spezialisierung auf forensische Methodik. Ein Nutzer beschreibt ein Szenario oder Problem. Analysieren Sie die Eingabe auf Vollständigkeit für eine wissenschaftlich fundierte Untersuchung.
ANALYSIEREN SIE DIESE FORENSISCHEN KATEGORIEN:
1. **Incident Context**: Was ist passiert? Welche Angriffsvektoren oder technischen Probleme liegen vor?
2. **Affected Systems**: Welche spezifischen Technologien/Plattformen sind betroffen? (Windows/Linux/ICS/SCADA/Mobile/Cloud/Network Infrastructure)
3. **Available Evidence**: Welche forensischen Datenquellen stehen zur Verfügung? (RAM-Dumps, Disk-Images, Log-Files, Network-Captures, Registry-Hives)
4. **Investigation Objectives**: Was soll erreicht werden? (IOC-Extraktion, Timeline-Rekonstruktion, Attribution, Impact-Assessment)
5. **Timeline Constraints**: Wie zeitkritisch ist die Untersuchung?
6. **Legal & Compliance**: Rechtliche Anforderungen, Chain of Custody, Compliance-Rahmen (DSGVO, sector-specific regulations)
7. **Technical Constraints**: Verfügbare Ressourcen, Skills, Infrastrukturbeschränkungen
WENN die Beschreibung alle kritischen forensischen Aspekte abdeckt: Geben Sie eine leere Liste [] zurück.
WENN wichtige Details fehlen: Formulieren Sie 2-3 präzise Fragen, die die kritischsten Lücken für eine wissenschaftlich fundierte Analyse schließen.
QUALITÄTSKRITERIEN FÜR FRAGEN:
- Forensisch spezifisch, nicht allgemein (NICHT: "Mehr Details?")
- Methodisch relevant (NICHT: "Wann passierte das?")
- Priorisiert nach Auswirkung auf die Untersuchungsqualität
- Die Frage soll maximal 20 Wörter umfassen
ANTWORTFORMAT (NUR JSON, KEIN ZUSÄTZLICHER TEXT):
[
"spezifische Frage 1?",
"spezifische Frage 2?",
"spezifische Frage 3?"
]
NUTZER-EINGABE:
${input}
`.trim();
} }
/**
* Handler
*/
export const POST: APIRoute = async ({ request }) => { export const POST: APIRoute = async ({ request }) => {
try { try {
const authResult = await withAPIAuth(request, 'ai'); const auth = await withAPIAuth(request, 'ai');
if (!authResult.authenticated) { if (!auth.authenticated) return createAuthErrorResponse();
return createAuthErrorResponse(); const userId = auth.userId;
}
const userId = authResult.userId;
if (!checkRateLimit(userId)) { if (!checkRateLimit(userId)) {
return apiError.rateLimit('Enhancement rate limit exceeded'); return apiError.rateLimit('Enhancement rate limit exceeded');
@ -100,61 +85,38 @@ export const POST: APIRoute = async ({ request }) => {
const body = await request.json(); const body = await request.json();
const { input } = body; const { input } = body;
if (!input || typeof input !== 'string' || input.length < 40) { if (!input || typeof input !== 'string' || input.length < INPUT_MIN_CHARS) {
return apiError.badRequest('Input too short for enhancement (minimum 40 characters)'); return apiError.badRequest(`Input too short for enhancement (minimum ${INPUT_MIN_CHARS} characters)`);
} }
const sanitizedInput = sanitizeInput(input); const sanitizedInput = sanitizeInput(input);
if (sanitizedInput.length < 40) { if (sanitizedInput.length < INPUT_MIN_CHARS) {
return apiError.badRequest('Input too short after sanitization'); return apiError.badRequest('Input too short after sanitization');
} }
const systemPrompt = createEnhancementPrompt(sanitizedInput); const taskId = `enhance_${userId}_${Date.now()}_${Math.random().toString(36).slice(2, 6)}`;
const taskId = `enhance_${userId}_${Date.now()}_${Math.random().toString(36).substr(2, 4)}`; const questionsPrompt = getPrompt('enhancementQuestions', sanitizedInput);
console.log(`[ENHANCE-API] Processing enhancement request for user: ${userId}`); console.log(`[ENHANCE-API] Processing enhancement request for user: ${userId}`);
const aiResponse = await enqueueApiCall(() => const aiResponse = await enqueueApiCall(
aiService.callAI(systemPrompt, { () => aiService.callAI(questionsPrompt, { temperature: AI_TEMPERATURE }),
temperature: 0.7 taskId
}), taskId); );
if (!aiResponse.content) { if (!aiResponse?.content) {
return apiServerError.unavailable('No enhancement response'); return apiServerError.unavailable('No enhancement response');
} }
let questions; let parsed: unknown = JSONParser.safeParseJSON(stripJsonFences(aiResponse.content), null);
try {
const cleanedContent = aiResponse.content
.replace(/^```json\s*/i, '')
.replace(/\s*```\s*$/, '')
.trim();
questions = JSONParser.safeParseJSON(cleanedContent, []); let questions: string[] = Array.isArray(parsed) ? parsed : [];
questions = questions
if (!Array.isArray(questions)) { .filter(q => typeof q === 'string')
throw new Error('Response is not an array'); .map(q => q.trim())
} .filter(q => q.endsWith('?'))
.filter(q => q.length >= Q_MIN_LEN && q.length <= Q_MAX_LEN)
questions = questions .slice(0, Q_MAX_COUNT);
.filter(q => typeof q === 'string' && q.length > 20 && q.length < 200)
.filter(q => q.includes('?'))
.filter(q => {
const forensicsTerms = ['forensisch', 'log', 'dump', 'image', 'artefakt', 'evidence', 'incident', 'system', 'netzwerk', 'zeitraum', 'verfügbar'];
const lowerQ = q.toLowerCase();
return forensicsTerms.some(term => lowerQ.includes(term));
})
.map(q => q.trim())
.slice(0, 3);
if (questions.length === 0) {
questions = [];
}
} catch (error) {
console.error('[ENHANCE-API] Failed to parse enhancement response:', aiResponse.content);
questions = [];
}
console.log(`[ENHANCE-API] User: ${userId}, Questions generated: ${questions.length}, Input length: ${sanitizedInput.length}`); console.log(`[ENHANCE-API] User: ${userId}, Questions generated: ${questions.length}, Input length: ${sanitizedInput.length}`);
@ -168,8 +130,8 @@ export const POST: APIRoute = async ({ request }) => {
headers: { 'Content-Type': 'application/json' } headers: { 'Content-Type': 'application/json' }
}); });
} catch (error) { } catch (err) {
console.error('[ENHANCE-API] Enhancement error:', error); console.error('[ENHANCE-API] Enhancement error:', err);
return apiServerError.internal('Enhancement processing failed'); return apiServerError.internal('Enhancement processing failed');
} }
}; };

View File

@ -470,13 +470,11 @@ class AIPipeline {
pipelineStart: number, pipelineStart: number,
toolsDataHash: string toolsDataHash: string
): Promise<{ completed: number; failed: number }> { ): Promise<{ completed: number; failed: number }> {
// Evaluate ALL candidates handed over by the embeddings pre-filter.
const candidates = context.filteredData.tools || []; const candidates = context.filteredData.tools || [];
if (!Array.isArray(candidates) || candidates.length === 0) { if (!Array.isArray(candidates) || candidates.length === 0) {
return { completed: completedTasks, failed: failedTasks }; return { completed: completedTasks, failed: failedTasks };
} }
// Evaluate every candidate (no slicing here)
for (let i = 0; i < candidates.length; i++) { for (let i = 0; i < candidates.length; i++) {
const evaluationResult = await this.evaluateSpecificTool(context, candidates[i], i + 1, pipelineStart, toolsDataHash); const evaluationResult = await this.evaluateSpecificTool(context, candidates[i], i + 1, pipelineStart, toolsDataHash);
if (evaluationResult.success) completedTasks++; else failedTasks++; if (evaluationResult.success) completedTasks++; else failedTasks++;
@ -484,15 +482,12 @@ class AIPipeline {
await this.delay(this.config.microTaskDelay); 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) { if (Array.isArray(context.selectedTools) && context.selectedTools.length > 0) {
context.selectedTools.sort((a: any, b: any) => { context.selectedTools.sort((a: any, b: any) => {
const ar = typeof a.taskRelevance === 'number' ? a.taskRelevance : -1; const ar = typeof a.taskRelevance === 'number' ? a.taskRelevance : -1;
const br = typeof b.taskRelevance === 'number' ? b.taskRelevance : -1; const br = typeof b.taskRelevance === 'number' ? b.taskRelevance : -1;
if (br !== ar) return br - ar; if (br !== ar) return br - ar;
// tie-breakers without domain heuristics:
const aLen = (a.justification || '').length; const aLen = (a.justification || '').length;
const bLen = (b.justification || '').length; const bLen = (b.justification || '').length;
if (bLen !== aLen) return bLen - aLen; if (bLen !== aLen) return bLen - aLen;
@ -502,7 +497,6 @@ class AIPipeline {
return aRank - bRank; return aRank - bRank;
}); });
// Keep top 3 only
context.selectedTools = context.selectedTools.slice(0, 3); context.selectedTools = context.selectedTools.slice(0, 3);
} }
@ -877,7 +871,6 @@ class AIPipeline {
): Promise<MicroTaskResult> { ): Promise<MicroTaskResult> {
const taskStart = Date.now(); const taskStart = Date.now();
// Build prompt WITHOUT any baseline score
const prompt = getPrompt('toolEvaluation', context.userQuery, tool, rank); const prompt = getPrompt('toolEvaluation', context.userQuery, tool, rank);
const result = await this.callMicroTaskAI(prompt, context, 'tool-evaluation'); const result = await this.callMicroTaskAI(prompt, context, 'tool-evaluation');
@ -885,16 +878,13 @@ class AIPipeline {
return result; return result;
} }
// Parse strictly; do NOT provide a default with a score.
const evaluation = JSONParser.safeParseJSON(result.content, null); 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) const aiProvided = evaluation && typeof evaluation.taskRelevance === 'number' && Number.isFinite(evaluation.taskRelevance)
? Math.round(evaluation.taskRelevance) ? Math.round(evaluation.taskRelevance)
: null; : null;
if (aiProvided === null) { if (aiProvided === null) {
// Log the malformed output but avoid injecting a synthetic score.
auditService.addAIDecision( auditService.addAIDecision(
'tool-evaluation', 'tool-evaluation',
prompt, prompt,
@ -920,7 +910,6 @@ class AIPipeline {
const moderatedTaskRelevance = this.moderateTaskRelevance(aiProvided); const moderatedTaskRelevance = this.moderateTaskRelevance(aiProvided);
const priority = this.derivePriorityFromScore(moderatedTaskRelevance); const priority = this.derivePriorityFromScore(moderatedTaskRelevance);
// Keep original fields if present; coerce to strings/arrays safely.
const detailed_explanation = String(evaluation?.detailed_explanation || '').trim(); const detailed_explanation = String(evaluation?.detailed_explanation || '').trim();
const implementation_approach = String(evaluation?.implementation_approach || '').trim(); const implementation_approach = String(evaluation?.implementation_approach || '').trim();
const pros = Array.isArray(evaluation?.pros) ? evaluation.pros : []; const pros = Array.isArray(evaluation?.pros) ? evaluation.pros : [];

View File

@ -194,18 +194,15 @@ class ToolSelector {
const softwareWithFullData = candidateSoftware.map(this.createToolData); const softwareWithFullData = candidateSoftware.map(this.createToolData);
const conceptsWithFullData = candidateConcepts.map(this.createConceptData); const conceptsWithFullData = candidateConcepts.map(this.createConceptData);
// Embeddings are always ON → only use embedding limits
const maxTools = Math.min(this.config.embeddingSelectionLimit, candidateTools.length); const maxTools = Math.min(this.config.embeddingSelectionLimit, candidateTools.length);
const maxConcepts = Math.min(this.config.embeddingConceptsLimit, candidateConcepts.length); const maxConcepts = Math.min(this.config.embeddingConceptsLimit, candidateConcepts.length);
// Respect ratios first, then fill the remaining capacity
const methodRatio = Math.max(0, Math.min(1, this.config.methodSelectionRatio)); const methodRatio = Math.max(0, Math.min(1, this.config.methodSelectionRatio));
const softwareRatio = Math.max(0, Math.min(1, this.config.softwareSelectionRatio)); const softwareRatio = Math.max(0, Math.min(1, this.config.softwareSelectionRatio));
let methodLimit = Math.round(maxTools * methodRatio); let methodLimit = Math.round(maxTools * methodRatio);
let softwareLimit = Math.round(maxTools * softwareRatio); let softwareLimit = Math.round(maxTools * softwareRatio);
// If rounded sum exceeds maxTools, scale down proportionally
if (methodLimit + softwareLimit > maxTools) { if (methodLimit + softwareLimit > maxTools) {
const scale = maxTools / (methodLimit + softwareLimit); const scale = maxTools / (methodLimit + softwareLimit);
methodLimit = Math.floor(methodLimit * scale); methodLimit = Math.floor(methodLimit * scale);
@ -217,7 +214,6 @@ class ToolSelector {
const toolsToSend: any[] = [...methodsPrimary, ...softwarePrimary]; const toolsToSend: any[] = [...methodsPrimary, ...softwarePrimary];
// Fill any remaining capacity from whichever pool still has candidates
let mIdx = methodsPrimary.length; let mIdx = methodsPrimary.length;
let sIdx = softwarePrimary.length; let sIdx = softwarePrimary.length;