fix(drafting): Drafting-Engine auf prod reparieren — RAG via ai-sdk + OVH-LLM-Kaskade
CI / python-lint (push) Has been skipped
CI / nodejs-lint (push) Has been skipped
CI / iace-gt-coverage (push) Has been skipped
CI / test-python-backend (push) Has been skipped
CI / test-python-document-crawler (push) Has been skipped
CI / test-python-dsms-gateway (push) Has been skipped
CI / detect-changes (push) Successful in 7s
CI / branch-name (push) Has been skipped
CI / guardrail-integrity (push) Has been skipped
CI / secret-scan (push) Has been skipped
CI / dep-audit (push) Has been skipped
CI / sbom-scan (push) Has been skipped
CI / build-sha-integrity (push) Successful in 5s
CI / validate-canonical-controls (push) Successful in 4s
CI / loc-budget (push) Successful in 17s
CI / go-lint (push) Has been skipped
CI / nodejs-build (push) Successful in 3m2s
CI / test-go (push) Has been skipped

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>
This commit is contained in:
Benjamin Admin
2026-06-19 10:02:06 +02:00
parent cd3e0b15ad
commit 90a70c8404
9 changed files with 398 additions and 203 deletions
@@ -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> {