fix(use-case-compiler): batch LLM calls + increase proxy timeout

Single LLM calls per MC caused 2min+ timeouts. Now batches up to 10
MCs in one prompt with 20s timeout. LLM failure falls through to
deterministic derivation gracefully. Proxy timeout increased to 60s.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

admin-compliance/app/api/sdk/v1/use-case/[[...path]]/route.ts

@@ -33,7 +33,7 @@ async function proxyRequest(
     const fetchOptions: RequestInit = {
       method,
       headers,
-      signal: AbortSignal.timeout(30000),
+      signal: AbortSignal.timeout(60000),
     }
 
     if (['POST', 'PUT', 'PATCH'].includes(method)) {

ai-compliance-sdk/internal/usecase/compiler_llm.go

@@ -23,63 +23,83 @@ func NewLLMQuestionGenerator(registry *llm.ProviderRegistry) *LLMQuestionGenerat
 
 // llmQuestion is the JSON structure we expect from the LLM.
 type llmQuestion struct {
+	MCName       string   `json:"mc_name"`
 	Question     string   `json:"question"`
 	PassCriteria []string `json:"pass_criteria"`
 	FailCriteria []string `json:"fail_criteria"`
 	Severity     string   `json:"severity"`
 }
 
-// GenerateQuestions generates questions for a list of MCs using the LLM.
+// maxLLMMCs limits how many MCs we send to the LLM in one batch.
+const maxLLMMCs = 10
+
+// GenerateQuestions generates questions for MCs using a single batched LLM call.
 func (g *LLMQuestionGenerator) GenerateQuestions(mcs []MCInfo, regulations []string) ([]Question, error) {
 	if g.registry == nil {
 		return nil, fmt.Errorf("no LLM provider configured")
 	}
 
-	ctx, cancel := context.WithTimeout(context.Background(), 120*time.Second)
+	// Limit batch size
+	batch := mcs
+	if len(batch) > maxLLMMCs {
+		batch = batch[:maxLLMMCs]
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 20*time.Second)
 	defer cancel()
 
+	prompt := buildBatchPrompt(batch, regulations)
+
+	resp, err := g.registry.Chat(ctx, &llm.ChatRequest{
+		Messages: []llm.Message{
+			{Role: "system", Content: systemPrompt},
+			{Role: "user", Content: prompt},
+		},
+		Temperature: 0.3,
+		MaxTokens:   2000,
+	})
+	if err != nil {
+		return nil, fmt.Errorf("LLM call failed: %w", err)
+	}
+
+	parsed := parseLLMResponse(resp.Message.Content)
+	if len(parsed) == 0 {
+		return nil, fmt.Errorf("LLM returned no valid questions")
+	}
+
+	// Map parsed questions back to MCs
+	mcByName := make(map[string]MCInfo)
+	for _, mc := range batch {
+		mcByName[mc.CanonicalName] = mc
+	}
+
 	var questions []Question
-	qNum := 1
-
-	for _, mc := range mcs {
-		prompt := buildPrompt(mc, regulations)
-
-		resp, err := g.registry.Chat(ctx, &llm.ChatRequest{
-			Messages: []llm.Message{
-				{Role: "system", Content: systemPrompt},
-				{Role: "user", Content: prompt},
-			},
-			Temperature: 0.3,
-			MaxTokens:   500,
-		})
-		if err != nil {
-			// Fallback to deterministic generation
-			questions = append(questions, GenerateFromMC(mc)...)
-			qNum += len(GenerateFromMC(mc))
-			continue
-		}
-
-		parsed := parseLLMResponse(resp.Message.Content)
-		for _, lq := range parsed {
-			q := Question{
-				ID:           fmt.Sprintf("Q%d", qNum),
-				MCID:         mc.MasterControlID,
-				MCName:       mc.CanonicalName,
-				Text:         lq.Question,
-				QuestionType: "yes_no",
-				Severity:     normalizeSeverity(lq.Severity),
-				Regulation:   mc.RegSource,
-				PassCriteria: lq.PassCriteria,
-				FailCriteria: lq.FailCriteria,
+	for _, lq := range parsed {
+		mc, ok := mcByName[lq.MCName]
+		if !ok {
+			// Try fuzzy match
+			for name, m := range mcByName {
+				if strings.Contains(lq.MCName, name) || strings.Contains(name, lq.MCName) {
+					mc = m
+					ok = true
+					break
+				}
 			}
-			questions = append(questions, q)
-			qNum++
 		}
 
-		// Cap total questions
-		if qNum > 50 {
-			break
+		q := Question{
+			Text:         lq.Question,
+			QuestionType: "yes_no",
+			Severity:     normalizeSeverity(lq.Severity),
+			PassCriteria: lq.PassCriteria,
+			FailCriteria: lq.FailCriteria,
 		}
+		if ok {
+			q.MCID = mc.MasterControlID
+			q.MCName = mc.CanonicalName
+			q.Regulation = mc.RegSource
+		}
+		questions = append(questions, q)
 	}
 
 	return questions, nil
@@ -88,12 +108,30 @@ func (g *LLMQuestionGenerator) GenerateQuestions(mcs []MCInfo, regulations []str
 const systemPrompt = `Du bist ein Compliance-Experte. Generiere praezise Prueffragen fuer Compliance-Audits.
 
 Antworte NUR mit einem JSON-Array. Jedes Element hat:
+- "mc_name": Der canonical_name des Master Controls (exakt wie im Input)
 - "question": Eine klare Ja/Nein-Frage auf Deutsch
 - "pass_criteria": Array mit 1-2 Kriterien fuer "bestanden"
 - "fail_criteria": Array mit 1-2 Kriterien fuer "nicht bestanden"
 - "severity": "HIGH", "MEDIUM" oder "LOW"
 
-Keine Erklaerungen, nur das JSON-Array.`
+Generiere 1 Frage pro Master Control. Keine Erklaerungen, nur das JSON-Array.`
+
+func buildBatchPrompt(mcs []MCInfo, regulations []string) string {
+	regStr := strings.Join(regulations, ", ")
+
+	var sb strings.Builder
+	sb.WriteString(fmt.Sprintf("Regulierungen: %s\n\nMaster Controls:\n", regStr))
+
+	for i, mc := range mcs {
+		readable := strings.ReplaceAll(mc.CanonicalName, "_", " ")
+		sb.WriteString(fmt.Sprintf("%d. mc_name=%q (%d Controls, Quelle: %s)\n",
+			i+1, mc.CanonicalName, mc.TotalControls, mc.RegSource))
+		_ = readable
+	}
+
+	sb.WriteString("\nGeneriere je 1 Prueffrage pro Master Control.")
+	return sb.String()
+}
 
 func buildPrompt(mc MCInfo, regulations []string) string {
 	readable := strings.ReplaceAll(mc.CanonicalName, "_", " ")