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:
1 parent
d41da670fc
commit
7026fc4fed
6 files changed
+589
-203
No files matched your search
@@ -1,13 +1,13 @@
|
||||
"""
|
||||
FILE: app/core/ingestion/ingestion_validation.py
|
||||
DESCRIPTION: WP-15b semantische Validierung von Kanten gegen den LocalBatchCache.
|
||||
WP-25a: Integration der Mixture of Experts (MoE) Profil-Steuerung.
|
||||
VERSION: 2.13.0 (WP-25a: MoE & Profile Support)
|
||||
WP-25b: Umstellung auf Lazy-Prompt-Orchestration (prompt_key + variables).
|
||||
VERSION: 2.14.0 (WP-25b: Lazy Prompt Integration)
|
||||
STATUS: Active
|
||||
FIX:
|
||||
- Umstellung auf generate_raw_response mit profile_name="ingest_validator".
|
||||
- Automatische Nutzung der Fallback-Kaskade (Cloud -> Lokal) via LLMService.
|
||||
- Erhalt der sparsamen LLM-Nutzung (Validierung nur für Kandidaten-Pool).
|
||||
- WP-25b: Entfernung manueller Prompt-Formatierung zur Unterstützung modell-spezifischer Prompts.
|
||||
- WP-25b: Umstellung auf generate_raw_response mit prompt_key="edge_validation".
|
||||
- WP-25a: Voller Erhalt der MoE-Profilsteuerung und Fallback-Kaskade via LLMService.
|
||||
"""
|
||||
import logging
|
||||
from typing import Dict, Any, Optional
|
||||
@@ -27,8 +27,8 @@ async def validate_edge_candidate(
|
||||
profile_name: str = "ingest_validator"
|
||||
) -> bool:
|
||||
"""
|
||||
WP-15b: Validiert einen Kandidaten semantisch gegen das Ziel im Cache.
|
||||
Nutzt das MoE-Profil 'ingest_validator' für deterministische YES/NO Prüfungen.
|
||||
WP-15b/25b: Validiert einen Kandidaten semantisch gegen das Ziel im Cache.
|
||||
Nutzt Lazy-Prompt-Loading zur Unterstützung modell-spezifischer Validierungs-Templates.
|
||||
"""
|
||||
target_id = edge.get("to")
|
||||
target_ctx = batch_cache.get(target_id)
|
||||
@@ -44,27 +44,25 @@ async def validate_edge_candidate(
|
||||
logger.info(f"ℹ️ [VALIDATION SKIP] No context for '{target_id}' - allowing link.")
|
||||
return True
|
||||
|
||||
# Prompt-Abruf (Nutzt Provider-String als Fallback-Key für die prompts.yaml)
|
||||
template = llm_service.get_prompt("edge_validation", provider)
|
||||
|
||||
try:
|
||||
logger.info(f"⚖️ [VALIDATING] Relation '{edge.get('kind')}' -> '{target_id}' (Profile: {profile_name})...")
|
||||
prompt = template.format(
|
||||
chunk_text=chunk_text[:1500],
|
||||
target_title=target_ctx.title,
|
||||
target_summary=target_ctx.summary,
|
||||
edge_kind=edge.get("kind", "related_to")
|
||||
)
|
||||
|
||||
# WP-25a: Profilbasierter Aufruf (Delegiert Fallbacks an den Service)
|
||||
# Nutzt ingest_validator (Cloud Mistral/Gemini -> Local Phi3:mini Kaskade)
|
||||
# WP-25b: Lazy-Prompt Aufruf.
|
||||
# Wir übergeben keine formatierte Nachricht mehr, sondern Key und Daten-Dict.
|
||||
# Das manuelle 'template = llm_service.get_prompt(...)' entfällt hier.
|
||||
raw_response = await llm_service.generate_raw_response(
|
||||
prompt,
|
||||
prompt_key="edge_validation",
|
||||
variables={
|
||||
"chunk_text": chunk_text[:1500],
|
||||
"target_title": target_ctx.title,
|
||||
"target_summary": target_ctx.summary,
|
||||
"edge_kind": edge.get("kind", "related_to")
|
||||
},
|
||||
priority="background",
|
||||
profile_name=profile_name
|
||||
)
|
||||
|
||||
# WP-14 Fix: Zusätzliche Bereinigung zur Sicherstellung der Interpretierbarkeit
|
||||
# WP-14 Fix: Bereinigung zur Sicherstellung der Interpretierbarkeit
|
||||
response = clean_llm_text(raw_response)
|
||||
|
||||
# Semantische Prüfung des Ergebnisses
|
||||
@@ -75,6 +73,7 @@ async def validate_edge_candidate(
|
||||
else:
|
||||
logger.info(f"🚫 [REJECTED] Relation to '{target_id}' irrelevant for this chunk.")
|
||||
return is_valid
|
||||
|
||||
except Exception as e:
|
||||
logger.warning(f"⚠️ Validation error for {target_id} using {profile_name}: {e}")
|
||||
# Im Zweifel (Timeout/Fehler) erlauben wir die Kante, um Datenverlust zu vermeiden
|
||||
|
||||
@@ -1,16 +1,15 @@
|
||||
"""
|
||||
FILE: app/core/retrieval/decision_engine.py
|
||||
DESCRIPTION: Der Agentic Orchestrator für MindNet (WP-25a Edition).
|
||||
DESCRIPTION: Der Agentic Orchestrator für MindNet (WP-25b Edition).
|
||||
Realisiert Multi-Stream Retrieval, Intent-basiertes Routing
|
||||
und die neue Pre-Synthesis Kompression (Module A).
|
||||
VERSION: 1.2.1 (WP-25a: Profile-Driven Orchestration & Optimized Cascade)
|
||||
und die neue Lazy-Prompt Orchestrierung (Module A & B).
|
||||
VERSION: 1.3.0 (WP-25b: Lazy Prompt Orchestration)
|
||||
STATUS: Active
|
||||
FIX:
|
||||
- WP-25a: Volle Integration der Profil-Kaskade (Delegation an LLMService v3.5.2).
|
||||
- WP-25a: Dynamische Nutzung des 'router_profile' für die Intent-Erkennung.
|
||||
- WP-25a: Parallelisierte Kompression überlanger Wissens-Streams.
|
||||
- WP-25b: Umstellung auf Lazy-Loading (Übergabe von prompt_key + variables).
|
||||
- WP-25b: Entfernung lokaler String-Formatierung zur Ermöglichung modell-spezifischer Prompts.
|
||||
- WP-25a: Volle Integration der Profil-Kaskade via LLMService v3.5.5.
|
||||
- WP-25: Beibehaltung von Stream-Tracing und Pre-Initialization Robustness.
|
||||
- CLEANUP: Entfernung redundanter Fallback-Blocks (jetzt im LLMService).
|
||||
"""
|
||||
import asyncio
|
||||
import logging
|
||||
@@ -70,29 +69,27 @@ class DecisionEngine:
|
||||
if not strategy:
|
||||
return "Entschuldigung, meine Wissensbasis ist aktuell nicht konfiguriert."
|
||||
|
||||
# 2. Multi-Stream Retrieval & Pre-Synthesis (Parallel Tasks)
|
||||
# WP-25a: Diese Methode übernimmt nun auch die Kompression.
|
||||
# 2. Multi-Stream Retrieval & Pre-Synthesis (Parallel Tasks inkl. Kompression)
|
||||
stream_results = await self._execute_parallel_streams(strategy, query)
|
||||
|
||||
# 3. Finale Synthese
|
||||
return await self._generate_final_answer(strategy_key, strategy, query, stream_results)
|
||||
|
||||
async def _determine_strategy(self, query: str) -> str:
|
||||
"""Nutzt den LLM-Router zur Wahl der Such-Strategie via router_profile."""
|
||||
"""WP-25b: Nutzt den LLM-Router via Lazy-Loading prompt_key."""
|
||||
settings_cfg = self.config.get("settings", {})
|
||||
prompt_key = settings_cfg.get("router_prompt_key", "intent_router_v1")
|
||||
# WP-25a: Nutzt das spezialisierte Profil für das Routing
|
||||
router_profile = settings_cfg.get("router_profile")
|
||||
|
||||
router_prompt_template = self.llm_service.get_prompt(prompt_key)
|
||||
if not router_prompt_template:
|
||||
return "FACT_WHAT"
|
||||
|
||||
full_prompt = router_prompt_template.format(query=query)
|
||||
try:
|
||||
# Der LLMService übernimmt hier über das Profil bereits die Fallback-Kaskade
|
||||
# WP-25b: Keine manuelle Formatierung mehr. Wir übergeben nur Key und Variablen.
|
||||
# Der LLMService wählt den passenden Prompt für das router_profile Modell.
|
||||
response = await self.llm_service.generate_raw_response(
|
||||
full_prompt, max_retries=1, priority="realtime", profile_name=router_profile
|
||||
prompt_key=prompt_key,
|
||||
variables={"query": query},
|
||||
max_retries=1,
|
||||
priority="realtime",
|
||||
profile_name=router_profile
|
||||
)
|
||||
return str(response).strip().upper()
|
||||
except Exception as e:
|
||||
@@ -102,7 +99,7 @@ class DecisionEngine:
|
||||
async def _execute_parallel_streams(self, strategy: Dict, query: str) -> Dict[str, str]:
|
||||
"""
|
||||
Führt Such-Streams aus und komprimiert überlange Ergebnisse (Pre-Synthesis).
|
||||
WP-25a: MoE-Profile werden für die Kompression berücksichtigt.
|
||||
WP-25b: Unterstützt Lazy-Compression über Experten-Profile.
|
||||
"""
|
||||
stream_keys = strategy.get("use_streams", [])
|
||||
library = self.config.get("streams_library", {})
|
||||
@@ -116,7 +113,7 @@ class DecisionEngine:
|
||||
active_streams.append(key)
|
||||
retrieval_tasks.append(self._run_single_stream(key, stream_cfg, query))
|
||||
|
||||
# Ergebnisse sammeln (Exceptions werden als Objekte zurückgegeben)
|
||||
# Ergebnisse sammeln
|
||||
retrieval_results = await asyncio.gather(*retrieval_tasks, return_exceptions=True)
|
||||
|
||||
# Phase 2: Formatierung und optionale Kompression
|
||||
@@ -137,38 +134,32 @@ class DecisionEngine:
|
||||
threshold = stream_cfg.get("compression_threshold", 4000)
|
||||
|
||||
if len(formatted_context) > threshold:
|
||||
logger.info(f"⚙️ [WP-25a] Compressing stream '{name}' ({len(formatted_context)} chars)...")
|
||||
logger.info(f"⚙️ [WP-25b] Triggering Lazy-Compression for stream '{name}'...")
|
||||
comp_profile = stream_cfg.get("compression_profile")
|
||||
final_stream_tasks.append(
|
||||
self._compress_stream_content(name, formatted_context, query, comp_profile)
|
||||
)
|
||||
else:
|
||||
# Direkt-Übernahme als Coroutine für gather()
|
||||
async def _direct(c=formatted_context): return c
|
||||
final_stream_tasks.append(_direct())
|
||||
|
||||
# Finale Inhalte (evtl. komprimiert) parallel fertigstellen
|
||||
# Finale Inhalte parallel fertigstellen
|
||||
final_contents = await asyncio.gather(*final_stream_tasks)
|
||||
|
||||
return dict(zip(active_streams, final_contents))
|
||||
|
||||
async def _compress_stream_content(self, stream_name: str, content: str, query: str, profile: Optional[str]) -> str:
|
||||
"""
|
||||
WP-25a Module A: Inhaltsverdichtung via Experten-Modell.
|
||||
"""
|
||||
compression_prompt = (
|
||||
f"Du bist ein Wissens-Analyst. Reduziere den folgenden Wissens-Stream '{stream_name}' "
|
||||
f"auf die Informationen, die für die Beantwortung der Frage '{query}' absolut notwendig sind.\n\n"
|
||||
f"BEIBEHALTEN: Harte Fakten, Projektnamen, konkrete Werte und Quellenangaben.\n"
|
||||
f"ENTFERNEN: Redundante Einleitungen, Füllwörter und irrelevante Details.\n\n"
|
||||
f"STREAM-INHALT:\n{content}\n\n"
|
||||
f"KOMPRIMIERTE ANALYSE:"
|
||||
)
|
||||
|
||||
"""WP-25b Module A: Inhaltsverdichtung via Lazy-Loading 'compression_template'."""
|
||||
try:
|
||||
# WP-25b: Wir übergeben den Auftrag an den LLMService.
|
||||
# Das Modell-spezifische Template wird erst beim Call aufgelöst.
|
||||
summary = await self.llm_service.generate_raw_response(
|
||||
compression_prompt,
|
||||
profile_name=profile, # WP-25a: MoE Support
|
||||
prompt_key="compression_template",
|
||||
variables={
|
||||
"stream_name": stream_name,
|
||||
"content": content,
|
||||
"query": query
|
||||
},
|
||||
profile_name=profile,
|
||||
priority="background",
|
||||
max_retries=1
|
||||
)
|
||||
@@ -191,24 +182,19 @@ class DecisionEngine:
|
||||
)
|
||||
|
||||
response = await self.retriever.search(request)
|
||||
|
||||
# WP-25: STREAM-TRACING
|
||||
for hit in response.results:
|
||||
hit.stream_origin = name
|
||||
|
||||
return response
|
||||
|
||||
def _format_stream_context(self, response: QueryResponse) -> str:
|
||||
"""Wandelt QueryHits in einen formatierten Kontext-String um."""
|
||||
if not response.results:
|
||||
return "Keine spezifischen Informationen in diesem Stream gefunden."
|
||||
|
||||
lines = []
|
||||
for i, hit in enumerate(response.results, 1):
|
||||
source = hit.source.get("path", "Unbekannt")
|
||||
content = hit.source.get("text", "").strip()
|
||||
lines.append(f"[{i}] QUELLE: {source}\nINHALT: {content}")
|
||||
|
||||
return "\n\n".join(lines)
|
||||
|
||||
async def _generate_final_answer(
|
||||
@@ -218,12 +204,9 @@ class DecisionEngine:
|
||||
query: str,
|
||||
stream_results: Dict[str, str]
|
||||
) -> str:
|
||||
"""Führt die finale Synthese basierend auf dem Strategie-Profil durch."""
|
||||
# WP-25a: Nutzt das llm_profile der Strategie
|
||||
"""WP-25b: Finale Synthese via Lazy-Prompt 'rag_template'."""
|
||||
profile = strategy.get("llm_profile")
|
||||
template_key = strategy.get("prompt_template", "rag_template")
|
||||
|
||||
template = self.llm_service.get_prompt(template_key)
|
||||
system_prompt = self.llm_service.get_prompt("system_prompt")
|
||||
|
||||
# WP-25 ROBUSTNESS: Pre-Initialization
|
||||
@@ -232,28 +215,15 @@ class DecisionEngine:
|
||||
template_vars.update(stream_results)
|
||||
template_vars["query"] = query
|
||||
|
||||
prepend = strategy.get("prepend_instruction", "")
|
||||
|
||||
# WP-25b: Wir reichen die Variablen direkt an den Service weiter.
|
||||
# Formatierung erfolgt erst nach Profil-Auflösung (Gemini vs. Llama vs. Phi3).
|
||||
try:
|
||||
final_prompt = template.format(**template_vars)
|
||||
if prepend:
|
||||
final_prompt = f"{prepend}\n\n{final_prompt}"
|
||||
|
||||
# WP-25a: MoE Call mit automatisierter Kaskade im LLMService
|
||||
# (Frühere manuelle Fallback-Blocks wurden entfernt, da v3.5.2 dies intern löst)
|
||||
response = await self.llm_service.generate_raw_response(
|
||||
final_prompt, system=system_prompt, profile_name=profile, priority="realtime"
|
||||
)
|
||||
|
||||
return response
|
||||
|
||||
except KeyError as e:
|
||||
logger.error(f"Template Variable mismatch in '{template_key}': Missing {e}")
|
||||
fallback_context = "\n\n".join([v for v in stream_results.values() if v])
|
||||
# WP-25a FIX: Nutzt auch im Fallback das Strategie-Profil für Konsistenz
|
||||
return await self.llm_service.generate_raw_response(
|
||||
f"Beantworte: {query}\n\nKontext:\n{fallback_context}",
|
||||
system=system_prompt, priority="realtime", profile_name=profile
|
||||
prompt_key=template_key,
|
||||
variables=template_vars,
|
||||
system=system_prompt,
|
||||
profile_name=profile,
|
||||
priority="realtime"
|
||||
)
|
||||
except Exception as e:
|
||||
logger.error(f"Final Synthesis failed: {e}")
|
||||
|
||||
+27
-25
@@ -1,15 +1,14 @@
|
||||
"""
|
||||
FILE: app/routers/chat.py
|
||||
DESCRIPTION: Haupt-Chat-Interface (WP-25a Agentic Edition).
|
||||
Kombiniert die spezialisierte Interview-Logik und Keyword-Erkennung
|
||||
mit der neuen MoE-Orchestrierung und Pre-Synthesis Kompression.
|
||||
VERSION: 3.0.4 (WP-25a: Optimized MoE & Cascade Delegation)
|
||||
DESCRIPTION: Haupt-Chat-Interface (WP-25b Edition).
|
||||
Kombiniert die spezialisierte Interview-Logik mit der neuen
|
||||
Lazy-Prompt-Orchestration und MoE-Synthese.
|
||||
VERSION: 3.0.5 (WP-25b: Lazy Prompt Integration)
|
||||
STATUS: Active
|
||||
FIX:
|
||||
- WP-25a: Delegation der Fallback-Kaskade an den LLMService (v3.5.2).
|
||||
- WP-25a: Nutzung der zentralisierten Stream-Kompression der DecisionEngine (v1.2.1).
|
||||
- WP-25a: Konsistente Nutzung von MoE-Profilen für Interview- und RAG-Modus.
|
||||
- 100% Erhalt der v3.0.2 Logik (Interview, Schema-Resolution, FastPaths).
|
||||
- WP-25b: Umstellung des Interview-Modus auf Lazy-Prompt (prompt_key + variables).
|
||||
- WP-25b: Delegation der RAG-Phase an die Engine v1.3.0 für konsistente MoE-Steuerung.
|
||||
- WP-25a: Voller Erhalt der v3.0.2 Logik (Interview, Schema-Resolution, FastPaths).
|
||||
"""
|
||||
|
||||
from fastapi import APIRouter, HTTPException, Depends
|
||||
@@ -146,7 +145,7 @@ def _collect_all_hits(stream_responses: Dict[str, Any]) -> List[QueryHit]:
|
||||
all_hits = []
|
||||
seen_node_ids = set()
|
||||
for _, response in stream_responses.items():
|
||||
# In v3.0.4 sammeln wir die hits aus den QueryResponse Objekten
|
||||
# Sammeln der Hits aus den QueryResponse Objekten
|
||||
if hasattr(response, 'results'):
|
||||
for hit in response.results:
|
||||
if hit.node_id not in seen_node_ids:
|
||||
@@ -166,8 +165,7 @@ async def chat_endpoint(
|
||||
):
|
||||
start_time = time.time()
|
||||
query_id = str(uuid.uuid4())
|
||||
settings = get_settings()
|
||||
logger.info(f"🚀 [WP-25a] Chat request [{query_id}]: {request.message[:50]}...")
|
||||
logger.info(f"🚀 [WP-25b] Chat request [{query_id}]: {request.message[:50]}...")
|
||||
|
||||
try:
|
||||
# 1. Intent Detection
|
||||
@@ -180,7 +178,7 @@ async def chat_endpoint(
|
||||
sources_hits = []
|
||||
answer_text = ""
|
||||
|
||||
# 2. INTERVIEW MODE (Bitgenaue WP-07 Logik)
|
||||
# 2. INTERVIEW MODE (WP-25b Lazy-Prompt Logik)
|
||||
if intent == "INTERVIEW":
|
||||
target_type = _detect_target_type(request.message, strategy.get("schemas", {}))
|
||||
types_cfg = get_types_config()
|
||||
@@ -194,23 +192,27 @@ async def chat_endpoint(
|
||||
fields_list = fallback.get("fields", []) if isinstance(fallback, dict) else (fallback or [])
|
||||
|
||||
fields_str = "\n- " + "\n- ".join(fields_list)
|
||||
template = llm.get_prompt(strategy.get("prompt_template", "interview_template"))
|
||||
template_key = strategy.get("prompt_template", "interview_template")
|
||||
|
||||
final_prompt = template.replace("{query}", request.message) \
|
||||
.replace("{target_type}", target_type) \
|
||||
.replace("{schema_fields}", fields_str)
|
||||
|
||||
# WP-25a: MoE Call (Kaskade erfolgt intern im LLMService)
|
||||
# WP-25b: Lazy Loading Call
|
||||
# Wir übergeben nur Key und Variablen. Das System formatiert passend zum Modell.
|
||||
answer_text = await llm.generate_raw_response(
|
||||
final_prompt, system=llm.get_prompt("system_prompt"),
|
||||
priority="realtime", profile_name="compression_fast", max_retries=0
|
||||
prompt_key=template_key,
|
||||
variables={
|
||||
"query": request.message,
|
||||
"target_type": target_type,
|
||||
"schema_fields": fields_str
|
||||
},
|
||||
system=llm.get_prompt("system_prompt"),
|
||||
priority="realtime",
|
||||
profile_name="compression_fast",
|
||||
max_retries=0
|
||||
)
|
||||
sources_hits = []
|
||||
|
||||
# 3. RAG MODE (Optimierte MoE Orchestrierung)
|
||||
# 3. RAG MODE (WP-25b Delegation an Engine v1.3.0)
|
||||
else:
|
||||
# Phase A & B: Retrieval & Kompression (Delegation an Engine v1.2.1)
|
||||
# Diese Methode gibt bereits die (evtl. komprimierten) Kontext-Strings zurück.
|
||||
# Phase A & B: Retrieval & Kompression (Delegiert an Engine v1.3.0)
|
||||
formatted_context_map = await engine._execute_parallel_streams(strategy, request.message)
|
||||
|
||||
# Erfassung der Quellen für das Tracing
|
||||
@@ -232,7 +234,7 @@ async def chat_endpoint(
|
||||
|
||||
sources_hits = _collect_all_hits(raw_stream_map)
|
||||
|
||||
# Phase C: Finale MoE Synthese
|
||||
# Phase C: Finale MoE Synthese (Delegiert an Engine v1.3.0)
|
||||
answer_text = await engine._generate_final_answer(
|
||||
intent, strategy, request.message, formatted_context_map
|
||||
)
|
||||
@@ -243,7 +245,7 @@ async def chat_endpoint(
|
||||
try:
|
||||
log_search(
|
||||
query_id=query_id, query_text=request.message, results=sources_hits,
|
||||
mode=f"wp25a_{intent.lower()}", metadata={"strategy": intent, "source": intent_source}
|
||||
mode=f"wp25b_{intent.lower()}", metadata={"strategy": intent, "source": intent_source}
|
||||
)
|
||||
except: pass
|
||||
|
||||
|
||||
+69
-81
@@ -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):
|
||||
|
||||
Reference in new issue
Block a user