Update Decision Engine and related components for WP-25a: Bump version to 1.2.0, enhance multi-stream retrieval with pre-synthesis compression, and integrate Mixture of Experts (MoE) profile support. Refactor chat interface to utilize new compression logic and llm_profiles for improved synthesis. Maintain compatibility with existing methods and ensure robust error handling across services.

This commit is contained in:
Lars committed 2026-01-02 07:04:43 +01:00
1 parent 3d2f3d12d9
commit d0eae8e43c
5 files changed
+299 -200

No files matched your search

+94 -90
View File
@@ -1,16 +1,14 @@
"""
FILE: app/services/llm_service.py
DESCRIPTION: Hybrid-Client für Ollama, Google GenAI (Gemini) und OpenRouter.
Verwaltet provider-spezifische Prompts und Background-Last.
WP-20: Optimiertes Fallback-Management zum Schutz von Cloud-Quoten.
WP-22/JSON: Optionales JSON-Schema + strict (für OpenRouter).
WP-25: Integration der DecisionEngine für Agentic Multi-Stream RAG.
VERSION: 3.4.2 (WP-25: Ingest-Stability Patch)
WP-25a: Implementierung der Mixture of Experts (MoE) Profil-Steuerung.
VERSION: 3.5.0 (WP-25a: MoE & Profile Orchestration)
STATUS: Active
FIX:
- Ingest-Stability: Entfernung des <5-Zeichen Guards (ermöglicht YES/NO Validierungen).
- OpenRouter-Fix: Sicherung gegen leere 'choices' zur Vermeidung von JSON-Errors.
- Erhalt der vollständigen v3.3.9 Logik für Rate-Limits, Retries und Background-Tasks.
- WP-25a: Profilbasiertes Routing via llm_profiles.yaml.
- WP-25a: Unterstützung individueller Temperaturen pro Experten-Profil.
- WP-25: Beibehaltung der Ingest-Stability (kein Schwellenwert für YES/NO).
- WP-25: Erhalt der vollständigen v3.4.2 Resilienz-Logik.
"""
import httpx
import yaml
@@ -19,28 +17,28 @@ import asyncio
import json
from google import genai
from google.genai import types
from openai import AsyncOpenAI # Für OpenRouter (OpenAI-kompatibel)
from openai import AsyncOpenAI
from pathlib import Path
from typing import Optional, Dict, Any, Literal
from app.config import get_settings
# ENTSCHEIDENDER FIX: Import der neutralen Bereinigungs-Logik (WP-14)
# Import der neutralen Bereinigungs-Logik
from app.core.registry import clean_llm_text
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-25: Lazy Initialization der DecisionEngine zur Vermeidung von Circular Imports
# WP-25a: Zentrale Experten-Profile laden
self.profiles = self._load_llm_profiles()
self._decision_engine = None
# Initialisiere Semaphore einmalig auf Klassen-Ebene
if LLMService._background_semaphore is None:
limit = getattr(self.settings, "BACKGROUND_LIMIT", 2)
logger.info(f"🚦 LLMService: Initializing Background Semaphore with limit: {limit}")
@@ -52,10 +50,9 @@ class LLMService:
timeout=httpx.Timeout(self.settings.LLM_TIMEOUT)
)
# 2. Google GenAI Client (Modern SDK)
# 2. Google GenAI Client
self.google_client = None
if self.settings.GOOGLE_API_KEY:
# FIX: Wir erzwingen api_version 'v1' für höhere Stabilität bei 2.5er Modellen.
self.google_client = genai.Client(
api_key=self.settings.GOOGLE_API_KEY,
http_options={'api_version': 'v1'}
@@ -68,24 +65,20 @@ class LLMService:
self.openrouter_client = AsyncOpenAI(
base_url="https://openrouter.ai/api/v1",
api_key=self.settings.OPENROUTER_API_KEY,
# Strikter Timeout für OpenRouter Free-Tier zur Vermeidung von Hangs.
timeout=45.0
)
logger.info("🛰️ LLMService: OpenRouter Integration active.")
@property
def decision_engine(self):
"""Lazy Initialization der Decision Engine (WP-25)."""
if self._decision_engine is None:
from app.core.retrieval.decision_engine import DecisionEngine
self._decision_engine = DecisionEngine()
return self._decision_engine
def _load_prompts(self) -> dict:
"""Lädt die Prompt-Konfiguration aus der YAML-Datei."""
path = Path(self.settings.PROMPTS_PATH)
if not path.exists():
logger.error(f"❌ Prompts file not found at {path}")
return {}
try:
with open(path, "r", encoding="utf-8") as f:
@@ -94,21 +87,28 @@ class LLMService:
logger.error(f"❌ Failed to load prompts: {e}")
return {}
def _load_llm_profiles(self) -> dict:
"""WP-25a: Lädt die zentralen MoE-Profile aus der llm_profiles.yaml."""
# Wir nutzen den in settings oder decision_engine definierten Pfad
path_str = getattr(self.settings, "LLM_PROFILES_PATH", "config/llm_profiles.yaml")
path = Path(path_str)
if not path.exists():
logger.warning(f"⚠️ LLM Profiles file not found at {path}. System will use .env defaults.")
return {}
try:
with open(path, "r", encoding="utf-8") as f:
data = yaml.safe_load(f) or {}
return data.get("profiles", {})
except Exception as e:
logger.error(f"❌ Failed to load llm_profiles.yaml: {e}")
return {}
def get_prompt(self, key: str, provider: str = None) -> str:
"""
Hole provider-spezifisches Template mit intelligenter Text-Kaskade.
Kaskade: Gewählter Provider -> Gemini -> Ollama.
"""
active_provider = provider or self.settings.MINDNET_LLM_PROVIDER
data = self.prompts.get(key, "")
if isinstance(data, dict):
val = data.get(active_provider, data.get("gemini", data.get("ollama", "")))
if isinstance(val, dict):
logger.warning(f"⚠️ [LLMService] Nested dictionary detected for key '{key}'. Using first entry.")
val = next(iter(val.values()), "") if val else ""
return str(val)
return str(data)
async def generate_raw_response(
@@ -123,34 +123,48 @@ class LLMService:
model_override: Optional[str] = None,
json_schema: Optional[Dict[str, Any]] = None,
json_schema_name: str = "mindnet_json",
strict_json_schema: bool = True
strict_json_schema: bool = True,
profile_name: Optional[str] = None # WP-25a
) -> str:
"""
Haupteinstiegspunkt für LLM-Anfragen.
WP-25 FIX: Schwellenwert entfernt, um kurze Ingest-Validierungen (YES/NO) zu unterstützen.
Haupteinstiegspunkt für LLM-Anfragen mit Profil-Unterstützung.
"""
target_provider = provider or self.settings.MINDNET_LLM_PROVIDER
target_provider = provider
target_model = model_override
target_temp = None
# WP-25a: Profil-Auflösung (Provider, Modell, Temperatur)
if profile_name and self.profiles:
profile = self.profiles.get(profile_name)
if profile:
target_provider = profile.get("provider", target_provider)
target_model = profile.get("model", target_model)
target_temp = profile.get("temperature")
logger.debug(f"🎭 MoE Call: Profil '{profile_name}' -> {target_provider}")
# Fallback auf Standard-Provider falls nichts übergeben/definiert wurde
if not target_provider:
target_provider = self.settings.MINDNET_LLM_PROVIDER
if priority == "background":
async with LLMService._background_semaphore:
res = await self._dispatch(
target_provider, prompt, system, force_json,
max_retries, base_delay, model_override,
json_schema, json_schema_name, strict_json_schema
max_retries, base_delay, target_model,
json_schema, json_schema_name, strict_json_schema, target_temp
)
else:
res = await self._dispatch(
target_provider, prompt, system, force_json,
max_retries, base_delay, model_override,
json_schema, json_schema_name, strict_json_schema
max_retries, base_delay, target_model,
json_schema, json_schema_name, strict_json_schema, target_temp
)
# WP-25 FIX: Nur noch auf absolut leere Antwort prüfen (ermöglicht YES/NO Antworten).
# WP-25 Fix: Ingest-Stability (Ermöglicht YES/NO ohne Schwellenwert-Blockade)
if not res and target_provider != "ollama":
logger.warning(f"⚠️ [WP-25] Empty response from {target_provider}. Falling back to OLLAMA.")
res = await self._execute_ollama(prompt, system, force_json, max_retries, base_delay)
logger.warning(f"⚠️ [WP-25] Empty response from {target_provider}. Fallback to OLLAMA.")
res = await self._execute_ollama(prompt, system, force_json, max_retries, base_delay, target_temp)
# WP-14 Fix: Bereinige Text-Antworten vor Rückgabe
return clean_llm_text(res) if not force_json else res
async def _dispatch(
@@ -164,9 +178,10 @@ class LLMService:
model_override: Optional[str],
json_schema: Optional[Dict[str, Any]],
json_schema_name: str,
strict_json_schema: bool
strict_json_schema: bool,
temperature: Optional[float] = None # WP-25a
) -> str:
"""Routet die Anfrage mit intelligenter Rate-Limit Erkennung."""
"""Routet die Anfrage mit Rate-Limit Erkennung."""
rate_limit_attempts = 0
max_rate_retries = min(max_retries, getattr(self.settings, "LLM_RATE_LIMIT_RETRIES", 3))
wait_time = getattr(self.settings, "LLM_RATE_LIMIT_WAIT", 60.0)
@@ -175,43 +190,42 @@ class LLMService:
try:
if provider == "openrouter" and self.openrouter_client:
return await self._execute_openrouter(
prompt=prompt,
system=system,
force_json=force_json,
model_override=model_override,
json_schema=json_schema,
json_schema_name=json_schema_name,
strict_json_schema=strict_json_schema
prompt=prompt, system=system, force_json=force_json,
model_override=model_override, json_schema=json_schema,
json_schema_name=json_schema_name, strict_json_schema=strict_json_schema,
temperature=temperature
)
if provider == "gemini" and self.google_client:
return await self._execute_google(prompt, system, force_json, model_override)
return await self._execute_google(prompt, system, force_json, model_override, temperature)
return await self._execute_ollama(prompt, system, force_json, max_retries, base_delay)
return await self._execute_ollama(prompt, system, force_json, max_retries, base_delay, temperature)
except Exception as e:
err_str = str(e)
is_rate_limit = any(x in err_str for x in ["429", "RESOURCE_EXHAUSTED", "rate_limited", "Too Many Requests"])
if is_rate_limit and rate_limit_attempts < max_rate_retries:
if any(x in err_str for x in ["429", "RESOURCE_EXHAUSTED", "rate_limited"]):
rate_limit_attempts += 1
logger.warning(f"⏳ Rate Limit from {provider}. Attempt {rate_limit_attempts}. Waiting {wait_time}s...")
logger.warning(f"⏳ Rate Limit {provider}. Attempt {rate_limit_attempts}. Wait {wait_time}s.")
await asyncio.sleep(wait_time)
continue
if self.settings.LLM_FALLBACK_ENABLED and provider != "ollama":
logger.warning(f"🔄 Provider {provider} failed ({err_str}). Falling back to OLLAMA.")
return await self._execute_ollama(prompt, system, force_json, max_retries, base_delay)
return await self._execute_ollama(prompt, system, force_json, max_retries, base_delay, temperature)
raise e
async def _execute_google(self, prompt, system, force_json, model_override):
async def _execute_google(self, prompt, system, force_json, model_override, temperature):
model = model_override or self.settings.GEMINI_MODEL
clean_model = model.replace("models/", "")
config = types.GenerateContentConfig(
system_instruction=system,
response_mime_type="application/json" if force_json else "text/plain"
)
config_kwargs = {
"system_instruction": system,
"response_mime_type": "application/json" if force_json else "text/plain"
}
if temperature is not None:
config_kwargs["temperature"] = temperature
config = types.GenerateContentConfig(**config_kwargs)
response = await asyncio.wait_for(
asyncio.to_thread(
self.google_client.models.generate_content,
@@ -222,53 +236,47 @@ class LLMService:
return response.text.strip()
async def _execute_openrouter(
self,
prompt: str,
system: Optional[str],
force_json: bool,
model_override: Optional[str],
json_schema: Optional[Dict[str, Any]] = None,
json_schema_name: str = "mindnet_json",
strict_json_schema: bool = True
self, prompt: str, system: Optional[str], force_json: bool,
model_override: Optional[str], json_schema: Optional[Dict[str, Any]] = None,
json_schema_name: str = "mindnet_json", strict_json_schema: bool = True,
temperature: Optional[float] = None
) -> str:
"""OpenRouter API Integration. WP-25 FIX: Sicherung gegen leere 'choices'."""
model = model_override or self.settings.OPENROUTER_MODEL
messages = []
if system:
messages.append({"role": "system", "content": system})
if system: messages.append({"role": "system", "content": system})
messages.append({"role": "user", "content": prompt})
kwargs: Dict[str, Any] = {}
if temperature is not None:
kwargs["temperature"] = temperature
if force_json:
if json_schema:
kwargs["response_format"] = {
"type": "json_schema",
"json_schema": {
"name": json_schema_name, "strict": strict_json_schema, "schema": json_schema
}
"json_schema": {"name": json_schema_name, "strict": strict_json_schema, "schema": json_schema}
}
else:
kwargs["response_format"] = {"type": "json_object"}
response = await self.openrouter_client.chat.completions.create(
model=model,
messages=messages,
**kwargs
model=model, messages=messages, **kwargs
)
# WP-25 FIX: Sicherung gegen leere Antwort-Arrays
if not response.choices or len(response.choices) == 0:
logger.warning(f"🛰️ OpenRouter returned no choices for model {model}")
if not response.choices:
return ""
return response.choices[0].message.content.strip() if response.choices[0].message.content else ""
async def _execute_ollama(self, prompt, system, force_json, max_retries, base_delay):
async def _execute_ollama(self, prompt, system, force_json, max_retries, base_delay, temperature=None):
# WP-25a: Nutzt Profil-Temperatur oder Standard
effective_temp = temperature if temperature is not None else (0.1 if force_json else 0.7)
payload = {
"model": self.settings.LLM_MODEL,
"prompt": prompt,
"stream": False,
"options": {"temperature": 0.1 if force_json else 0.7, "num_ctx": 8192}
"options": {"temperature": effective_temp, "num_ctx": 8192}
}
if force_json: payload["format"] = "json"
if system: payload["system"] = system
@@ -281,15 +289,11 @@ class LLMService:
return res.json().get("response", "").strip()
except Exception as e:
attempt += 1
if attempt > max_retries:
logger.error(f"❌ Ollama request failed: {e}")
raise e
wait_time = base_delay * (2 ** (attempt - 1))
await asyncio.sleep(wait_time)
if attempt > max_retries: raise e
await asyncio.sleep(base_delay * (2 ** (attempt - 1)))
async def generate_rag_response(self, query: str, context_str: Optional[str] = None) -> str:
"""WP-25: Orchestrierung via DecisionEngine."""
logger.info(f"🚀 [WP-25] Chat Query: {query[:50]}...")
return await self.decision_engine.ask(query)
async def close(self):