komplette openrouter integration

This commit is contained in:
Lars committed 2025-12-23 15:55:06 +01:00
1 parent c60aba63a4
commit 36fb27edf0
4 files changed
+96 -53

No files matched your search

+61 -28
View File
@@ -1,8 +1,10 @@
"""
FILE: app/services/llm_service.py
DESCRIPTION: Hybrid-Client für Ollama, Google GenAI und OpenRouter.
DESCRIPTION: Hybrid-Client für Ollama, Google GenAI (Gemini) und OpenRouter.
Verwaltet provider-spezifische Prompts und Background-Last.
VERSION: 3.3.0 (Full SDK Integration)
WP-20: Optimiertes Fallback-Management zum Schutz von Cloud-Quoten.
VERSION: 3.3.1
STATUS: Active
"""
import httpx
import yaml
@@ -11,7 +13,7 @@ import asyncio
import json
from google import genai
from google.genai import types
from openai import AsyncOpenAI # Für OpenRouter
from openai import AsyncOpenAI # Für OpenRouter (OpenAI-kompatibel)
from pathlib import Path
from typing import Optional, Dict, Any, Literal
from app.config import get_settings
@@ -19,16 +21,17 @@ from app.config import get_settings
logger = logging.getLogger(__name__)
class LLMService:
# GLOBALER SEMAPHOR für Hintergrund-Last Steuerung (WP-06)
_background_semaphore = None
def __init__(self):
self.settings = get_settings()
self.prompts = self._load_prompts()
# WP-06: Semaphore-Initialisierung
# Initialisiere Semaphore einmalig auf Klassen-Ebene
if LLMService._background_semaphore is None:
limit = self.settings.BACKGROUND_LIMIT
logger.info(f"🚦 LLMService: Background Semaphore initialized with limit: {limit}")
limit = getattr(self.settings, "BACKGROUND_LIMIT", 2)
logger.info(f"🚦 LLMService: Initializing Background Semaphore with limit: {limit}")
LLMService._background_semaphore = asyncio.Semaphore(limit)
# 1. Lokaler Ollama Client
@@ -53,6 +56,7 @@ class LLMService:
logger.info("🛰️ LLMService: OpenRouter Integration active.")
def _load_prompts(self) -> dict:
"""Lädt die Prompt-Konfiguration aus der YAML-Datei."""
path = Path(self.settings.PROMPTS_PATH)
if not path.exists(): return {}
try:
@@ -62,11 +66,16 @@ class LLMService:
return {}
def get_prompt(self, key: str, provider: str = None) -> str:
"""Hole provider-spezifisches Template mit Fallback-Kaskade."""
"""
Hole provider-spezifisches Template mit intelligenter Text-Kaskade.
HINWEIS: Dies ist nur ein Text-Lookup und verbraucht kein API-Kontingent.
Kaskade: Gewählter Provider -> Gemini (Cloud-Stil) -> Ollama (Basis-Stil).
"""
active_provider = provider or self.settings.MINDNET_LLM_PROVIDER
data = self.prompts.get(key, "")
if isinstance(data, dict):
return data.get(active_provider, data.get("ollama", ""))
# Wir versuchen erst den Provider, dann Gemini (weil ähnlich leistungsfähig), dann Ollama
return data.get(active_provider, data.get("gemini", data.get("ollama", "")))
return str(data)
async def generate_raw_response(
@@ -76,35 +85,43 @@ class LLMService:
provider: Optional[str] = None,
model_override: Optional[str] = None
) -> str:
"""Einstiegspunkt mit Priority-Handling."""
"""Haupteinstiegspunkt für LLM-Anfragen mit Priorisierung."""
target_provider = provider or self.settings.MINDNET_LLM_PROVIDER
if priority == "background":
async with LLMService._background_semaphore:
return await self._dispatch(target_provider, prompt, system, force_json, max_retries, base_delay, model_override)
return await self._dispatch(target_provider, prompt, system, force_json, max_retries, base_delay, model_override)
async def _dispatch(self, provider, prompt, system, force_json, max_retries, base_delay, model_override):
"""Routet die Anfrage an den physikalischen API-Provider."""
try:
if provider == "openrouter" and self.openrouter_client:
return await self._execute_openrouter(prompt, system, force_json, model_override)
if provider == "gemini" and self.google_client:
return await self._execute_google(prompt, system, force_json, model_override)
# Default/Fallback zu Ollama
return await self._execute_ollama(prompt, system, force_json, max_retries, base_delay)
except Exception as e:
# QUOTEN-SCHUTZ: Wenn Cloud (OpenRouter/Gemini) fehlschlägt,
# gehen wir IMMER zu Ollama, niemals von OpenRouter zu Gemini.
if self.settings.LLM_FALLBACK_ENABLED and provider != "ollama":
logger.warning(f"🔄 Provider {provider} failed: {e}. Falling back to Ollama.")
logger.warning(f"🔄 Provider {provider} failed: {e}. Falling back to LOCAL OLLAMA to protect cloud quotas.")
return await self._execute_ollama(prompt, system, force_json, max_retries, base_delay)
raise e
async def _execute_google(self, prompt, system, force_json, model_override):
"""Native Google SDK Integration."""
"""Native Google SDK Integration (Gemini)."""
model = model_override or self.settings.GEMINI_MODEL
config = types.GenerateContentConfig(
system_instruction=system,
response_mime_type="application/json" if force_json else "text/plain"
)
# Synchroner SDK-Call in Thread auslagern
# SDK Call in Thread auslagern, da die Google API blocking sein kann
response = await asyncio.to_thread(
self.google_client.models.generate_content,
model=model, contents=prompt, config=config
@@ -112,10 +129,11 @@ class LLMService:
return response.text.strip()
async def _execute_openrouter(self, prompt, system, force_json, model_override):
"""OpenRouter (OpenAI-kompatibel)."""
model = model_override or self.settings.OPENROUTER_MODEL
"""OpenRouter API Integration (OpenAI-kompatibel)."""
model = model_override or getattr(self.settings, "OPENROUTER_MODEL", "google/gemma-2-9b-it:free")
messages = []
if system: messages.append({"role": "system", "content": system})
if system:
messages.append({"role": "system", "content": system})
messages.append({"role": "user", "content": prompt})
response = await self.openrouter_client.chat.completions.create(
@@ -126,10 +144,15 @@ class LLMService:
return response.choices[0].message.content.strip()
async def _execute_ollama(self, prompt, system, force_json, max_retries, base_delay):
"""Ollama mit exponentiellem Backoff."""
"""Lokaler Ollama Call mit exponentiellem Backoff."""
payload = {
"model": self.settings.LLM_MODEL, "prompt": prompt, "stream": False,
"options": {"temperature": 0.1 if force_json else 0.7, "num_ctx": 8192}
"model": self.settings.LLM_MODEL,
"prompt": prompt,
"stream": False,
"options": {
"temperature": 0.1 if force_json else 0.7,
"num_ctx": 8192
}
}
if force_json: payload["format"] = "json"
if system: payload["system"] = system
@@ -142,18 +165,28 @@ class LLMService:
return res.json().get("response", "").strip()
except Exception as e:
attempt += 1
if attempt > max_retries: raise e
wait = base_delay * (2 ** (attempt - 1))
logger.warning(f"⚠️ Ollama retry {attempt} in {wait}s...")
await asyncio.sleep(wait)
if attempt > max_retries:
logger.error(f"Ollama Error after {attempt} retries: {e}")
raise e
wait_time = base_delay * (2 ** (attempt - 1))
logger.warning(f"⚠️ Ollama attempt {attempt} failed. Retrying in {wait_time}s...")
await asyncio.sleep(wait_time)
async def generate_rag_response(self, query: str, context_str: str) -> str:
"""Vollständiger RAG-Wrapper."""
"""Vollständiges RAG Chat-Interface."""
provider = self.settings.MINDNET_LLM_PROVIDER
system = self.get_prompt("system_prompt", provider)
template = self.get_prompt("rag_template", provider)
final_prompt = template.format(context_str=context_str, query=query)
return await self.generate_raw_response(final_prompt, system=system, priority="realtime")
system_prompt = self.get_prompt("system_prompt", provider)
rag_template = self.get_prompt("rag_template", provider)
final_prompt = rag_template.format(context_str=context_str, query=query)
return await self.generate_raw_response(
final_prompt,
system=system_prompt,
priority="realtime"
)
async def close(self):
await self.ollama_client.aclose()
"""Schließt die HTTP-Verbindungen."""
if self.ollama_client:
await self.ollama_client.aclose()