breakpilot-compliance/admin-compliance/lib/sdk/drafting-engine/llm-cascade.ts

/**
 * Gemeinsame LLM-Kaskade fuer die Drafting-Engine.
 *
 * Reihenfolge: OVH/LiteLLM (gpt-oss-120b) zuerst — der prod-Chat-LLM; Ollama als
 * Dev-Fallback. Auf prod ist OLLAMA_URL embedding-only (kein Chat-Modell), daher
 * ist OVH dort der einzige funktionierende Pfad — genau wie beim Compliance-Advisor
 * (siehe lib/sdk/agents/advisor-llm). Das Backend nutzt OVH + JSON-Output bereits
 * erfolgreich auf prod (extract-datasheet), dieselbe Technik wird hier gespiegelt.
 */

const OLLAMA_URL = process.env.OLLAMA_URL || 'http://host.docker.internal:11434'
const OLLAMA_MODEL = process.env.COMPLIANCE_LLM_MODEL || 'qwen2.5vl:32b'
const OVH_URL = (process.env.OVH_LLM_URL || '').replace(/\/+$/, '')
const OVH_MODEL = process.env.OVH_LLM_MODEL || ''
const OVH_KEY = process.env.OVH_LLM_KEY || ''

export interface ChatMessage {
  role: string
  content: string
}

export interface CascadeOpts {
  json?: boolean
  maxTokens?: number
  temperature?: number
  timeoutMs?: number
}

export interface CascadeResult {
  content: string
  tokensUsed: number
  provider: 'ovh' | 'ollama'
}

// --- Stream-Parser (pure, testbar) ---

/** Text-Delta aus einer Ollama-NDJSON-Zeile (message.content). */
export function parseOllamaLine(line: string): string | null {
  const t = line.trim()
  if (!t) return null
  try {
    return JSON.parse(t)?.message?.content || null
  } catch {
    return null
  }
}

/** Text-Delta aus einer OpenAI/OVH-SSE-Zeile (choices[0].delta.content). */
export function parseSSELine(line: string): string | null {
  const t = line.trim()
  if (!t.startsWith('data:')) return null
  const payload = t.slice(5).trim()
  if (!payload || payload === '[DONE]') return null
  try {
    return JSON.parse(payload)?.choices?.[0]?.delta?.content || null
  } catch {
    return null
  }
}

// --- Non-Streaming (cascadeComplete) ---

async function ovhComplete(messages: ChatMessage[], o: CascadeOpts): Promise<CascadeResult | null> {
  if (!OVH_URL || !OVH_MODEL) return null
  try {
    const headers: Record<string, string> = { 'Content-Type': 'application/json' }
    if (OVH_KEY) headers['Authorization'] = `Bearer ${OVH_KEY}`
    const payload: Record<string, unknown> = {
      model: OVH_MODEL,
      messages,
      stream: false,
      temperature: o.temperature ?? 0.15,
      max_tokens: o.maxTokens ?? 4096,
    }
    if (o.json) payload.response_format = { type: 'json_object' }
    const r = await fetch(`${OVH_URL}/v1/chat/completions`, {
      method: 'POST',
      headers,
      body: JSON.stringify(payload),
      signal: AbortSignal.timeout(o.timeoutMs ?? 120000),
    })
    if (!r.ok) return null
    const d = await r.json()
    const content = d?.choices?.[0]?.message?.content || ''
    if (!content) return null
    const usage = d?.usage || {}
    return { content, tokensUsed: usage.completion_tokens ?? usage.total_tokens ?? 0, provider: 'ovh' }
  } catch {
    return null
  }
}

async function ollamaComplete(messages: ChatMessage[], o: CascadeOpts): Promise<CascadeResult | null> {
  try {
    const body: Record<string, unknown> = {
      model: OLLAMA_MODEL,
      messages,
      stream: false,
      think: false,
      options: { temperature: o.temperature ?? 0.15, num_predict: o.maxTokens ?? 4096, num_ctx: 8192 },
    }
    if (o.json) body.format = 'json'
    const r = await fetch(`${OLLAMA_URL}/api/chat`, {
      method: 'POST',
      headers: { 'Content-Type': 'application/json' },
      body: JSON.stringify(body),
      signal: AbortSignal.timeout(o.timeoutMs ?? 120000),
    })
    if (!r.ok) return null
    const d = await r.json()
    const content = d?.message?.content || ''
    if (!content) return null
    return { content, tokensUsed: d?.eval_count ?? 0, provider: 'ollama' }
  } catch {
    return null
  }
}

/**
 * Nicht-streamender LLM-Aufruf mit Kaskade. Liefert Inhalt + Token + Provider,
 * oder null wenn weder OVH noch Ollama antworten.
 */
export async function cascadeComplete(messages: ChatMessage[], opts: CascadeOpts = {}): Promise<CascadeResult | null> {
  return (await ovhComplete(messages, opts)) ?? (await ollamaComplete(messages, opts))
}

// --- Streaming (cascadeStream) ---

const encoder = new TextEncoder()

function textStream(upstream: Response, parseLine: (line: string) => string | null): ReadableStream<Uint8Array> {
  return new ReadableStream({
    async start(controller) {
      const reader = upstream.body!.getReader()
      const decoder = new TextDecoder()
      let buf = ''
      try {
        for (;;) {
          const { done, value } = await reader.read()
          if (done) break
          buf += decoder.decode(value, { stream: true })
          const lines = buf.split('\n')
          buf = lines.pop() || ''
          for (const line of lines) {
            const delta = parseLine(line)
            if (delta) controller.enqueue(encoder.encode(delta))
          }
        }
        const tail = parseLine(buf)
        if (tail) controller.enqueue(encoder.encode(tail))
      } finally {
        controller.close()
      }
    },
  })
}

async function ovhStream(messages: ChatMessage[], o: CascadeOpts): Promise<Response | null> {
  if (!OVH_URL || !OVH_MODEL) return null
  try {
    const headers: Record<string, string> = { 'Content-Type': 'application/json' }
    if (OVH_KEY) headers['Authorization'] = `Bearer ${OVH_KEY}`
    const r = await fetch(`${OVH_URL}/v1/chat/completions`, {
      method: 'POST',
      headers,
      body: JSON.stringify({
        model: OVH_MODEL,
        messages,
        stream: true,
        temperature: o.temperature ?? 0.3,
        max_tokens: o.maxTokens ?? 4096,
      }),
      signal: AbortSignal.timeout(o.timeoutMs ?? 120000),
    })
    return r.ok && r.body ? r : null
  } catch {
    return null
  }
}

async function ollamaStream(messages: ChatMessage[], o: CascadeOpts): Promise<Response | null> {
  try {
    const r = await fetch(`${OLLAMA_URL}/api/chat`, {
      method: 'POST',
      headers: { 'Content-Type': 'application/json' },
      body: JSON.stringify({
        model: OLLAMA_MODEL,
        messages,
        stream: true,
        think: false,
        options: { temperature: o.temperature ?? 0.3, num_predict: o.maxTokens ?? 4096, num_ctx: 8192 },
      }),
      signal: AbortSignal.timeout(o.timeoutMs ?? 120000),
    })
    return r.ok && r.body ? r : null
  } catch {
    return null
  }
}

/**
 * Streamender LLM-Aufruf mit Kaskade -> Plain-Text-Stream. null = kein LLM erreichbar.
 */
export async function cascadeStream(messages: ChatMessage[], opts: CascadeOpts = {}): Promise<ReadableStream<Uint8Array> | null> {
  const ovh = await ovhStream(messages, opts)
  if (ovh) return textStream(ovh, parseSSELine)
  const ollama = await ollamaStream(messages, opts)
  if (ollama) return textStream(ollama, parseOllamaLine)
  return null
}