Update components for WP-25b: Implement Lazy-Prompt-Orchestration across ingestion, decision engine, chat interface, and LLM service. Enhance prompt management with hierarchical model support and streamline response generation by removing manual formatting. Bump versions to reflect new features and optimizations.

This commit is contained in:
Lars committed 2026-01-02 20:43:31 +01:00
1 parent d41da670fc
commit 7026fc4fed
6 files changed
+589 -203

No files matched your search

+69 -81
View File
@@ -1,14 +1,14 @@
"""
FILE: app/services/llm_service.py
DESCRIPTION: Hybrid-Client für Ollama, Google GenAI (Gemini) und OpenRouter.
WP-25a: Implementierung der Mixture of Experts (MoE) Kaskaden-Steuerung.
VERSION: 3.5.2 (WP-25a: MoE & Fallback Cascade Support)
WP-25b: Implementierung der Lazy-Prompt-Orchestration (Modell-spezifisch).
VERSION: 3.5.5 (WP-25b: Prompt Orchestration & Full Resilience)
STATUS: Active
FIX:
- WP-25a: Implementierung der rekursiven Fallback-Kaskade via fallback_profile.
- WP-25a: Schutz gegen zirkuläre Profil-Referenzen (visited_profiles).
- WP-25a: Erweitertes Logging für Tracing der Experten-Entscheidungen.
- Erhalt der Ingest-Stability (WP-25) und des Rate-Limit-Managements.
- WP-25b: get_prompt() unterstützt Hierarchie: Model-ID -> Provider -> Default.
- WP-25b: generate_raw_response() unterstützt prompt_key + variables für Lazy-Formatting.
- WP-25a: Voller Erhalt der rekursiven Fallback-Kaskade und visited_profiles Schutz.
- WP-20: Restaurierung des internen Ollama-Retry-Loops für Hardware-Stabilität.
"""
import httpx
import yaml
@@ -33,10 +33,7 @@ class LLMService:
def __init__(self):
self.settings = get_settings()
self.prompts = self._load_prompts()
# WP-25a: Zentrale Experten-Profile laden
self.profiles = self._load_llm_profiles()
self._decision_engine = None
if LLMService._background_semaphore is None:
@@ -92,7 +89,7 @@ class LLMService:
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.")
logger.warning(f"⚠️ LLM Profiles file not found at {path}.")
return {}
try:
with open(path, "r", encoding="utf-8") as f:
@@ -102,17 +99,31 @@ class LLMService:
logger.error(f"❌ Failed to load llm_profiles.yaml: {e}")
return {}
def get_prompt(self, key: str, provider: str = None) -> str:
active_provider = provider or self.settings.MINDNET_LLM_PROVIDER
def get_prompt(self, key: str, model_id: str = None, provider: str = None) -> str:
"""
WP-25b: Hochpräziser Prompt-Lookup.
Hierarchie: Exakte Modell-ID -> Provider-Name -> Globaler Default.
"""
data = self.prompts.get(key, "")
if isinstance(data, dict):
val = data.get(active_provider, data.get("gemini", data.get("ollama", "")))
return str(val)
return str(data)
if not isinstance(data, dict):
return str(data)
# 1. Spezifischstes Match: Exakte Modell-ID (z.B. 'meta-llama/llama-3.3-70b-instruct:free')
if model_id and model_id in data:
return str(data[model_id])
# 2. Mittlere Ebene: Provider (z.B. 'ollama' oder 'openrouter')
if provider and provider in data:
return str(data[provider])
# 3. Fallback: Bekannte Keys oder Default aus prompts.yaml
return str(data.get("default", data.get("gemini", data.get("ollama", ""))))
async def generate_raw_response(
self,
prompt: str,
prompt: str = None,
prompt_key: str = None, # WP-25b: Lazy Loading Key
variables: dict = None, # WP-25b: Daten für Formatierung
system: str = None,
force_json: bool = False,
max_retries: int = 2,
@@ -126,16 +137,14 @@ class LLMService:
profile_name: Optional[str] = None,
visited_profiles: Optional[list] = None
) -> str:
"""
Haupteinstiegspunkt für LLM-Anfragen mit rekursiver Kaskaden-Logik.
"""
"""Haupteinstiegspunkt für LLM-Anfragen mit Lazy-Prompt Orchestrierung."""
visited_profiles = visited_profiles or []
target_provider = provider
target_model = model_override
target_temp = None
fallback_profile = None
# 1. Profil-Auflösung
# 1. Profil-Auflösung (Mixture of Experts)
if profile_name and self.profiles:
profile = self.profiles.get(profile_name)
if profile:
@@ -148,30 +157,39 @@ class LLMService:
else:
logger.warning(f"⚠️ Profil '{profile_name}' nicht in llm_profiles.yaml gefunden!")
# Fallback auf Standard-Provider falls nichts übergeben/definiert wurde
if not target_provider:
target_provider = self.settings.MINDNET_LLM_PROVIDER
logger.info(f"ℹ️ Kein Provider/Profil definiert. Nutze Default: {target_provider}")
# 2. Ausführung mit Fehler-Handling für Kaskade
# 2. WP-25b: Lazy Prompt Resolving
# Wir laden den Prompt erst JETZT, basierend auf dem gerade aktiven Modell.
current_prompt = prompt
if prompt_key:
template = self.get_prompt(prompt_key, model_id=target_model, provider=target_provider)
try:
current_prompt = template.format(**(variables or {}))
except Exception as e:
logger.error(f"❌ Prompt formatting failed for key '{prompt_key}': {e}")
current_prompt = template # Sicherheits-Fallback
# 3. Ausführung mit Fehler-Handling für Kaskade
try:
if priority == "background":
async with LLMService._background_semaphore:
res = await self._dispatch(
target_provider, prompt, system, force_json,
target_provider, current_prompt, system, force_json,
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,
target_provider, current_prompt, system, force_json,
max_retries, base_delay, target_model,
json_schema, json_schema_name, strict_json_schema, target_temp
)
# Check auf leere Cloud-Antworten (WP-25 Stability)
if not res and target_provider != "ollama":
logger.warning(f"⚠️ Empty response from {target_provider}. Triggering fallback chain.")
logger.warning(f"⚠️ Empty response from {target_provider}. Triggering fallback.")
raise ValueError(f"Empty response from {target_provider}")
return clean_llm_text(res) if not force_json else res
@@ -179,40 +197,33 @@ class LLMService:
except Exception as e:
logger.error(f"❌ Error during execution of profile '{profile_name}' ({target_provider}): {e}")
# 3. Kaskaden-Logik: Nächstes Profil in der Kette versuchen
# 4. WP-25b Kaskaden-Logik (Rekursiv mit Modell-spezifischem Re-Loading)
if fallback_profile and fallback_profile not in visited_profiles:
logger.info(f"🔄 Switching to fallback profile: '{fallback_profile}'")
return await self.generate_raw_response(
prompt=prompt, system=system, force_json=force_json,
prompt=prompt,
prompt_key=prompt_key,
variables=variables, # Ermöglicht neues Formatting für Fallback-Modell
system=system, force_json=force_json,
max_retries=max_retries, base_delay=base_delay,
priority=priority, provider=provider, model_override=model_override,
priority=priority, provider=None, model_override=None,
json_schema=json_schema, json_schema_name=json_schema_name,
strict_json_schema=strict_json_schema,
profile_name=fallback_profile,
visited_profiles=visited_profiles
)
# 4. Ultimativer Notanker: Falls alles fehlschlägt, direkt zu Ollama
# 5. Ultimativer Notanker: Falls alles fehlschlägt, direkt zu Ollama
if target_provider != "ollama" and self.settings.LLM_FALLBACK_ENABLED:
logger.warning(f"🚨 Kaskade erschöpft. Nutze finalen Ollama-Notanker.")
res = await self._execute_ollama(prompt, system, force_json, max_retries, base_delay)
res = await self._execute_ollama(current_prompt, system, force_json, max_retries, base_delay, target_temp, target_model)
return clean_llm_text(res) if not force_json else res
raise e
async def _dispatch(
self,
provider: str,
prompt: str,
system: Optional[str],
force_json: bool,
max_retries: int,
base_delay: float,
model_override: Optional[str],
json_schema: Optional[Dict[str, Any]],
json_schema_name: str,
strict_json_schema: bool,
temperature: Optional[float] = None
self, provider, prompt, system, force_json, max_retries, base_delay,
model_override, json_schema, json_schema_name, strict_json_schema, temperature
) -> str:
"""Routet die Anfrage an den spezifischen Provider-Executor."""
rate_limit_attempts = 0
@@ -232,23 +243,19 @@ class LLMService:
if provider == "gemini" and self.google_client:
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, temperature)
return await self._execute_ollama(prompt, system, force_json, max_retries, base_delay, temperature, model_override)
except Exception as e:
err_str = str(e)
# Rate-Limit Handling (429)
if any(x in err_str for x in ["429", "RESOURCE_EXHAUSTED", "rate_limited"]):
rate_limit_attempts += 1
logger.warning(f"⏳ Rate Limit {provider}. Attempt {rate_limit_attempts}. Wait {wait_time}s.")
await asyncio.sleep(wait_time)
continue
# Andere Fehler werden an generate_raw_response für die Kaskade gereicht
raise e
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/", "")
model = (model_override or self.settings.GEMINI_MODEL).replace("models/", "")
config_kwargs = {
"system_instruction": system,
"response_mime_type": "application/json" if force_json else "text/plain"
@@ -257,22 +264,13 @@ class LLMService:
config_kwargs["temperature"] = temperature
config = types.GenerateContentConfig(**config_kwargs)
response = await asyncio.wait_for(
asyncio.to_thread(
self.google_client.models.generate_content,
model=clean_model, contents=prompt, config=config
),
asyncio.to_thread(self.google_client.models.generate_content, model=model, contents=prompt, config=config),
timeout=45.0
)
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,
temperature: Optional[float] = None
) -> str:
async def _execute_openrouter(self, prompt, system, force_json, model_override, json_schema, json_schema_name, strict_json_schema, temperature) -> str:
model = model_override or self.settings.OPENROUTER_MODEL
logger.info(f"🛰️ OpenRouter Call: Model='{model}' | Temp={temperature}")
messages = []
@@ -280,35 +278,26 @@ class LLMService:
messages.append({"role": "user", "content": prompt})
kwargs: Dict[str, Any] = {}
if temperature is not None:
kwargs["temperature"] = temperature
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}
}
kwargs["response_format"] = {"type": "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
)
if not response.choices:
return ""
response = await self.openrouter_client.chat.completions.create(model=model, messages=messages, **kwargs)
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, temperature=None):
# Nutzt Profil-Temperatur oder strikte Defaults für lokale Hardware-Schonung
async def _execute_ollama(self, prompt, system, force_json, max_retries, base_delay, temperature=None, model_override=None):
# WP-20: Restaurierter Retry-Loop für lokale Hardware-Resilienz
effective_model = model_override or self.settings.LLM_MODEL
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,
"model": effective_model,
"prompt": prompt, "stream": False,
"options": {"temperature": effective_temp, "num_ctx": 8192}
}
if force_json: payload["format"] = "json"
@@ -323,12 +312,11 @@ class LLMService:
except Exception as e:
attempt += 1
if attempt > max_retries:
logger.error(f"❌ Ollama final failure after {attempt} attempts: {e}")
logger.error(f"❌ Ollama failure after {attempt} attempts: {e}")
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."""
return await self.decision_engine.ask(query)
async def close(self):