fix(drafting): Drafting-Engine auf prod reparieren — RAG via ai-sdk + OVH-LLM-Kaskade

Die Drafting-Engine (Dokument-Entwurf, v2-Pipeline, Validierung, Drafting-Chat,
Vendor-Vertragspruefung) war auf prod doppelt tot:
- RAG ueber bp-core-rag-service:8097 (existiert auf prod nicht)
- LLM ueber OLLAMA_URL/api/chat mit qwen2.5vl (prod = ollama-embed, kein Chat-Modell)

Fix (analog zum Compliance-Advisor):
- rag-query.ts -> ai-compliance-sdk /sdk/v1/rag/search (bge-m3, prod-erreichbar).
- Neue lib/sdk/drafting-engine/llm-cascade.ts: OVH/LiteLLM (gpt-oss-120b) zuerst,
  Ollama als Dev-Fallback; cascadeComplete (JSON) + cascadeStream. Das Backend nutzt
  OVH+JSON bereits erfolgreich auf prod (extract-datasheet).
- 5 Aufrufstellen (draft-helpers, draft-helpers-v2, validate, chat, vendor-review)
  auf die Kaskade umgestellt; keine direkten Ollama-Calls mehr.
- Tests: llm-cascade + rag-query aktualisiert.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

admin-compliance/app/api/sdk/drafting-engine/chat/route.ts

@@ -11,9 +11,7 @@ import { queryRAG } from '@/lib/sdk/drafting-engine/rag-query'
 import { DOCUMENT_RAG_CONFIG } from '@/lib/sdk/drafting-engine/rag-config'
 import { readSoulFile } from '@/lib/sdk/agents/soul-reader'
 import type { ScopeDocumentType } from '@/lib/sdk/compliance-scope-types'
-
-const OLLAMA_URL = process.env.OLLAMA_URL || 'http://host.docker.internal:11434'
-const LLM_MODEL = process.env.COMPLIANCE_LLM_MODEL || 'qwen2.5vl:32b'
+import { cascadeStream } from '@/lib/sdk/drafting-engine/llm-cascade'
 
 // Fallback SOUL prompt (used when .soul.md file is unavailable)
 const FALLBACK_DRAFTING_PROMPT = `# Drafting Agent - Compliance-Dokumententwurf
@@ -81,66 +79,20 @@ export async function POST(request: NextRequest) {
     ]
 
     // 4. Call LLM with streaming
-    const ollamaResponse = await fetch(`${OLLAMA_URL}/api/chat`, {
-      method: 'POST',
-      headers: { 'Content-Type': 'application/json' },
-      body: JSON.stringify({
-        model: LLM_MODEL,
-        messages,
-        stream: true,
-        think: false,
-        options: {
-          temperature: mode === 'draft' ? 0.2 : 0.3,
-          num_predict: mode === 'draft' ? 16384 : 8192,
-          num_ctx: 8192,
-        },
-      }),
-      signal: AbortSignal.timeout(120000),
+    // 4. LLM-Kaskade (OVH -> Ollama) -> Plain-Text-Stream
+    const stream = await cascadeStream(messages, {
+      temperature: mode === 'draft' ? 0.2 : 0.3,
+      maxTokens: mode === 'draft' ? 16384 : 8192,
+      timeoutMs: 120000,
     })
 
-    if (!ollamaResponse.ok) {
-      const errorText = await ollamaResponse.text()
-      console.error('LLM error:', ollamaResponse.status, errorText)
+    if (!stream) {
       return NextResponse.json(
-        { error: `LLM nicht erreichbar (Status ${ollamaResponse.status})` },
+        { error: 'LLM nicht erreichbar (weder OVH noch Ollama)' },
         { status: 502 }
       )
     }
 
-    // 5. Stream response back
-    const encoder = new TextEncoder()
-    const stream = new ReadableStream({
-      async start(controller) {
-        const reader = ollamaResponse.body!.getReader()
-        const decoder = new TextDecoder()
-
-        try {
-          while (true) {
-            const { done, value } = await reader.read()
-            if (done) break
-
-            const chunk = decoder.decode(value, { stream: true })
-            const lines = chunk.split('\n').filter((l) => l.trim())
-
-            for (const line of lines) {
-              try {
-                const json = JSON.parse(line)
-                if (json.message?.content) {
-                  controller.enqueue(encoder.encode(json.message.content))
-                }
-              } catch {
-                // Partial JSON, skip
-              }
-            }
-          }
-        } catch (error) {
-          console.error('Stream error:', error)
-        } finally {
-          controller.close()
-        }
-      },
-    })
-
     return new NextResponse(stream, {
       headers: {
         'Content-Type': 'text/plain; charset=utf-8',

admin-compliance/app/api/sdk/drafting-engine/draft/draft-helpers-v2.ts

@@ -17,6 +17,7 @@ import { executeRepairLoop, type ProseBlockOutput, type RepairAudit } from '@/li
 import { computeChecksumSync, type CacheKeyParams } from '@/lib/sdk/drafting-engine/cache'
 import { queryRAG } from '@/lib/sdk/drafting-engine/rag-query'
 import { DOCUMENT_RAG_CONFIG } from '@/lib/sdk/drafting-engine/rag-config'
+import { cascadeComplete } from '@/lib/sdk/drafting-engine/llm-cascade'
 import {
   constraintEnforcer,
   proseCache,
@@ -27,7 +28,6 @@ import {
   buildPromptForDocumentType,
 } from './draft-helpers'
 
-const OLLAMA_URL = process.env.OLLAMA_URL || 'http://host.docker.internal:11434'
 const LLM_MODEL = process.env.COMPLIANCE_LLM_MODEL || 'qwen2.5vl:32b'
 
 // ============================================================================
@@ -171,29 +171,15 @@ Keine neuen Fakten erfinden — nur das Profil wuerdigen.`
 }
 
 export async function callOllama(systemPrompt: string, userPrompt: string): Promise<string> {
-  const response = await fetch(`${OLLAMA_URL}/api/chat`, {
-    method: 'POST',
-    headers: { 'Content-Type': 'application/json' },
-    body: JSON.stringify({
-      model: LLM_MODEL,
-      messages: [
-        { role: 'system', content: systemPrompt },
-        { role: 'user', content: userPrompt },
-      ],
-      stream: false,
-      think: false,
-      options: { temperature: 0.15, num_predict: 4096, num_ctx: 8192 },
-      format: 'json',
-    }),
-    signal: AbortSignal.timeout(120000),
-  })
-
-  if (!response.ok) {
-    throw new Error(`Ollama error: ${response.status}`)
-  }
-
-  const result = await response.json()
-  return result.message?.content || ''
+  const llm = await cascadeComplete(
+    [
+      { role: 'system', content: systemPrompt },
+      { role: 'user', content: userPrompt },
+    ],
+    { json: true, temperature: 0.15, maxTokens: 8192, timeoutMs: 120000 },
+  )
+  if (!llm) throw new Error('LLM nicht erreichbar (weder OVH noch Ollama)')
+  return llm.content
 }
 
 export async function handleV2Draft(body: Record<string, unknown>): Promise<NextResponse> {

admin-compliance/app/api/sdk/drafting-engine/draft/draft-helpers.ts

@@ -17,9 +17,7 @@ import { ConstraintEnforcer } from '@/lib/sdk/drafting-engine/constraint-enforce
 import { ProseCacheManager } from '@/lib/sdk/drafting-engine/cache'
 import { queryRAG } from '@/lib/sdk/drafting-engine/rag-query'
 import { DOCUMENT_RAG_CONFIG } from '@/lib/sdk/drafting-engine/rag-config'
-
-const OLLAMA_URL = process.env.OLLAMA_URL || 'http://host.docker.internal:11434'
-const LLM_MODEL = process.env.COMPLIANCE_LLM_MODEL || 'qwen2.5vl:32b'
+import { cascadeComplete } from '@/lib/sdk/drafting-engine/llm-cascade'
 
 export const constraintEnforcer = new ConstraintEnforcer()
 export const proseCache = new ProseCacheManager({ maxEntries: 200, ttlHours: 24 })
@@ -105,29 +103,21 @@ export async function handleV1Draft(body: Record<string, unknown>): Promise<Next
     { role: 'user', content: draftPrompt },
   ]
 
-  const ollamaResponse = await fetch(`${OLLAMA_URL}/api/chat`, {
-    method: 'POST',
-    headers: { 'Content-Type': 'application/json' },
-    body: JSON.stringify({
-      model: LLM_MODEL,
-      messages,
-      stream: false,
-      think: false,
-      options: { temperature: 0.15, num_predict: 16384, num_ctx: 8192 },
-      format: 'json',
-    }),
-    signal: AbortSignal.timeout(180000),
+  const llm = await cascadeComplete(messages, {
+    json: true,
+    temperature: 0.15,
+    maxTokens: 16384,
+    timeoutMs: 180000,
   })
 
-  if (!ollamaResponse.ok) {
+  if (!llm) {
     return NextResponse.json(
-      { error: `LLM nicht erreichbar (Status ${ollamaResponse.status})` },
+      { error: 'LLM nicht erreichbar (weder OVH noch Ollama)' },
       { status: 502 }
     )
   }
 
-  const result = await ollamaResponse.json()
-  const content = result.message?.content || ''
+  const content = llm.content
 
   let sections: DraftSection[] = []
   try {
@@ -153,7 +143,7 @@ export async function handleV1Draft(body: Record<string, unknown>): Promise<Next
   return NextResponse.json({
     draft,
     constraintCheck,
-    tokensUsed: result.eval_count || 0,
+    tokensUsed: llm.tokensUsed,
   } satisfies DraftResponse)
 }
 

admin-compliance/app/api/sdk/drafting-engine/validate/route.ts

@@ -10,9 +10,7 @@ import { DOCUMENT_SCOPE_MATRIX_CORE, DOCUMENT_TYPE_LABELS, getDepthLevelNumeric
 import type { ScopeDocumentType, ComplianceDepthLevel } from '@/lib/sdk/compliance-scope-types'
 import type { ValidationContext, ValidationResult, ValidationFinding } from '@/lib/sdk/drafting-engine/types'
 import { buildCrossCheckPrompt } from '@/lib/sdk/drafting-engine/prompts/validate-cross-check'
-
-const OLLAMA_URL = process.env.OLLAMA_URL || 'http://host.docker.internal:11434'
-const LLM_MODEL = process.env.COMPLIANCE_LLM_MODEL || 'qwen2.5vl:32b'
+import { cascadeComplete } from '@/lib/sdk/drafting-engine/llm-cascade'
 
 /**
  * Anti-Fake-Evidence: Verbotene Formulierungen
@@ -244,30 +242,17 @@ export async function POST(request: NextRequest) {
         context: validationContext,
       })
 
-      const ollamaResponse = await fetch(`${OLLAMA_URL}/api/chat`, {
-        method: 'POST',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({
-          model: LLM_MODEL,
-          messages: [
-            {
-              role: 'system',
-              content: 'Du bist ein DSGVO-Compliance-Validator. Antworte NUR im JSON-Format.',
-            },
-            { role: 'user', content: crossCheckPrompt },
-          ],
-          stream: false,
-          think: false,
-          options: { temperature: 0.1, num_predict: 8192, num_ctx: 8192 },
-          format: 'json',
-        }),
-        signal: AbortSignal.timeout(120000),
-      })
+      const llm = await cascadeComplete(
+        [
+          { role: 'system', content: 'Du bist ein DSGVO-Compliance-Validator. Antworte NUR im JSON-Format.' },
+          { role: 'user', content: crossCheckPrompt },
+        ],
+        { json: true, temperature: 0.1, maxTokens: 8192, timeoutMs: 120000 },
+      )
 
-      if (ollamaResponse.ok) {
-        const result = await ollamaResponse.json()
+      if (llm) {
         try {
-          const parsed = JSON.parse(result.message?.content || '{}')
+          const parsed = JSON.parse(llm.content || '{}')
           llmFindings = [
             ...(parsed.errors || []),
             ...(parsed.warnings || []),

admin-compliance/app/api/sdk/v1/vendor-compliance/contracts/[id]/review/route.ts

@@ -6,9 +6,7 @@ import {
 } from '@/lib/sdk/vendor-compliance'
 import { queryRAG } from '@/lib/sdk/drafting-engine/rag-query'
 import { transformAnalysisResponse } from '@/lib/sdk/vendor-compliance/contract-review/analyzer'
-
-const OLLAMA_URL = process.env.OLLAMA_URL || 'http://host.docker.internal:11434'
-const LLM_MODEL = process.env.COMPLIANCE_LLM_MODEL || 'qwen2.5vl:32b'
+import { cascadeComplete } from '@/lib/sdk/drafting-engine/llm-cascade'
 
 /**
  * POST /api/sdk/v1/vendor-compliance/contracts/[id]/review
@@ -47,29 +45,19 @@ export async function POST(
         }
 
         // Call Ollama
-        const ollamaResponse = await fetch(`${OLLAMA_URL}/api/chat`, {
-          method: 'POST',
-          headers: { 'Content-Type': 'application/json' },
-          body: JSON.stringify({
-            model: LLM_MODEL,
-            messages: [
-              { role: 'system', content: systemPrompt },
-              { role: 'user', content: `Analysiere den folgenden Vertrag auf DSGVO-Konformitaet:\n\n${documentText}` },
-            ],
-            stream: false,
-            options: { temperature: 0.1, num_predict: 16384 },
-            format: 'json',
-          }),
-          signal: AbortSignal.timeout(180000),
-        })
+        const llm = await cascadeComplete(
+          [
+            { role: 'system', content: systemPrompt },
+            { role: 'user', content: `Analysiere den folgenden Vertrag auf DSGVO-Konformitaet:\n\n${documentText}` },
+          ],
+          { json: true, temperature: 0.1, maxTokens: 16384, timeoutMs: 180000 },
+        )
 
-        if (!ollamaResponse.ok) {
-          throw new Error(`LLM nicht erreichbar (Status ${ollamaResponse.status})`)
+        if (!llm) {
+          throw new Error('LLM nicht erreichbar (weder OVH noch Ollama)')
         }
 
-        const result = await ollamaResponse.json()
-        const content = result.message?.content || ''
-        const llmResponse = JSON.parse(content)
+        const llmResponse = JSON.parse(llm.content)
 
         // Transform LLM response to typed findings
         const analysisResult = transformAnalysisResponse(llmResponse, {