Update FastAPI application and related services for WP-25a: Enhance lifespan management with Mixture of Experts (MoE) integrity checks, improve logging and error handling in LLMService, and integrate profile-driven orchestration across components. Bump versions for main application, ingestion services, and LLM profiles to reflect new features and optimizations.
This commit is contained in:
1 parent
9a98093e70
commit
9b906bbabf
9 files changed
+408
-195
No files matched your search
+31
-58
@@ -3,14 +3,13 @@ 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: 3.0.3 (WP-25a: MoE & Compression Support - Full Release)
|
||||
VERSION: 3.0.4 (WP-25a: Optimized MoE & Cascade Delegation)
|
||||
STATUS: Active
|
||||
FIX:
|
||||
- 100% Wiederherstellung der v3.0.2 Logik (Interview Fallbacks, Schema-Resolution).
|
||||
- WP-25a: Integration der Stream-Kompression (Module A) in den RAG-Workflow.
|
||||
- WP-25a: Unterstützung der llm_profiles für spezialisierte Synthese (Module B).
|
||||
- Erhalt der Ollama Context-Throttling Parameter (WP-20) als finaler Schutz.
|
||||
- Beibehaltung der No-Retry Logik (max_retries=0) für Chat-Stabilität.
|
||||
- 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).
|
||||
"""
|
||||
|
||||
from fastapi import APIRouter, HTTPException, Depends
|
||||
@@ -31,13 +30,13 @@ from app.services.feedback_service import log_search
|
||||
router = APIRouter()
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
# --- EBENE 1: CONFIG LOADER & CACHING (Restauriert aus v3.0.2) ---
|
||||
# --- EBENE 1: CONFIG LOADER & CACHING (WP-25 Standard) ---
|
||||
|
||||
_DECISION_CONFIG_CACHE = None
|
||||
_TYPES_CONFIG_CACHE = None
|
||||
|
||||
def _load_decision_config() -> Dict[str, Any]:
|
||||
"""Lädt die Strategie-Konfiguration (Kompatibilität zu WP-25)."""
|
||||
"""Lädt die Strategie-Konfiguration."""
|
||||
settings = get_settings()
|
||||
path = Path(settings.DECISION_CONFIG_PATH)
|
||||
try:
|
||||
@@ -49,7 +48,7 @@ def _load_decision_config() -> Dict[str, Any]:
|
||||
return {"strategies": {}}
|
||||
|
||||
def _load_types_config() -> Dict[str, Any]:
|
||||
"""Lädt die types.yaml für die Typerkennung im Interview-Modus."""
|
||||
"""Lädt die types.yaml für die Typerkennung."""
|
||||
path = os.getenv("MINDNET_TYPES_FILE", "config/types.yaml")
|
||||
try:
|
||||
if os.path.exists(path):
|
||||
@@ -109,7 +108,7 @@ def _detect_target_type(message: str, configured_schemas: Dict[str, Any]) -> str
|
||||
return "default"
|
||||
|
||||
def _is_question(query: str) -> bool:
|
||||
"""Prüft, ob der Input eine Frage ist (W-Fragen Erkennung)."""
|
||||
"""Prüft, ob der Input eine Frage ist."""
|
||||
q = query.strip().lower()
|
||||
if "?" in q: return True
|
||||
starters = ["wer", "wie", "was", "wo", "wann", "warum", "weshalb", "wozu", "welche", "bist du"]
|
||||
@@ -136,17 +135,18 @@ async def _classify_intent(query: str, llm: LLMService) -> tuple[str, str]:
|
||||
if kw.lower() in query_lower:
|
||||
return "INTERVIEW", "Keyword (Interview)"
|
||||
|
||||
# 3. SLOW PATH: DecisionEngine LLM Router
|
||||
# 3. SLOW PATH: DecisionEngine LLM Router (MoE-gesteuert)
|
||||
intent = await llm.decision_engine._determine_strategy(query)
|
||||
return intent, "DecisionEngine (LLM)"
|
||||
|
||||
# --- EBENE 3: RETRIEVAL AGGREGATION ---
|
||||
|
||||
def _collect_all_hits(stream_responses: Dict[str, Any]) -> List[QueryHit]:
|
||||
"""Sammelt und dedupliziert Treffer aus allen parallelen Streams."""
|
||||
"""Sammelt deduplizierte Treffer aus allen Streams für das Tracing."""
|
||||
all_hits = []
|
||||
seen_node_ids = set()
|
||||
for _, response in stream_responses.items():
|
||||
# In v3.0.4 sammeln wir die hits aus den QueryResponse Objekten
|
||||
if hasattr(response, 'results'):
|
||||
for hit in response.results:
|
||||
if hit.node_id not in seen_node_ids:
|
||||
@@ -180,14 +180,14 @@ async def chat_endpoint(
|
||||
sources_hits = []
|
||||
answer_text = ""
|
||||
|
||||
# 2. INTERVIEW MODE (Kompatibilität zu v3.0.2)
|
||||
# 2. INTERVIEW MODE (Bitgenaue WP-07 Logik)
|
||||
if intent == "INTERVIEW":
|
||||
target_type = _detect_target_type(request.message, strategy.get("schemas", {}))
|
||||
types_cfg = get_types_config()
|
||||
type_def = types_cfg.get("types", {}).get(target_type, {})
|
||||
fields_list = type_def.get("schema", [])
|
||||
|
||||
# WP-07: RESTAURIERTE FALLBACK LOGIK (v3.0.2)
|
||||
# WP-07: Restaurierte Fallback Logik
|
||||
if not fields_list:
|
||||
configured_schemas = strategy.get("schemas", {})
|
||||
fallback = configured_schemas.get(target_type, configured_schemas.get("default", {}))
|
||||
@@ -200,73 +200,46 @@ async def chat_endpoint(
|
||||
.replace("{target_type}", target_type) \
|
||||
.replace("{schema_fields}", fields_str)
|
||||
|
||||
# WP-25a: Nutzt spezialisiertes Kompressions-Profil für Interviews
|
||||
# WP-25a: MoE Call (Kaskade erfolgt intern im LLMService)
|
||||
answer_text = await llm.generate_raw_response(
|
||||
final_prompt, system=llm.get_prompt("system_prompt"),
|
||||
priority="realtime", profile_name="compression_fast", max_retries=0
|
||||
)
|
||||
sources_hits = []
|
||||
|
||||
# 3. RAG MODE (WP-25a Multi-Stream + Pre-Synthesis)
|
||||
# 3. RAG MODE (Optimierte MoE Orchestrierung)
|
||||
else:
|
||||
# Phase A & B: Retrieval & Kompression (Delegation an Engine v1.2.1)
|
||||
# Diese Methode gibt bereits die (evtl. komprimierten) Kontext-Strings zurück.
|
||||
formatted_context_map = await engine._execute_parallel_streams(strategy, request.message)
|
||||
|
||||
# Erfassung der Quellen für das Tracing
|
||||
raw_stream_map = {}
|
||||
stream_keys = strategy.get("use_streams", [])
|
||||
library = engine.config.get("streams_library", {})
|
||||
|
||||
# Phase A: Retrieval
|
||||
tasks = []
|
||||
retrieval_tasks = []
|
||||
active_streams = []
|
||||
for key in stream_keys:
|
||||
stream_cfg = library.get(key)
|
||||
if stream_cfg:
|
||||
if key in library:
|
||||
active_streams.append(key)
|
||||
tasks.append(engine._run_single_stream(key, stream_cfg, request.message))
|
||||
|
||||
responses = await asyncio.gather(*tasks, return_exceptions=True)
|
||||
retrieval_tasks.append(engine._run_single_stream(key, library[key], request.message))
|
||||
|
||||
raw_stream_map = {}
|
||||
formatted_context_tasks = []
|
||||
max_chars = getattr(settings, "MAX_OLLAMA_CHARS", 10000)
|
||||
provider = strategy.get("preferred_provider") or settings.MINDNET_LLM_PROVIDER
|
||||
|
||||
# Phase B: Pre-Synthesis & Throttling
|
||||
responses = await asyncio.gather(*retrieval_tasks, return_exceptions=True)
|
||||
for name, res in zip(active_streams, responses):
|
||||
if not isinstance(res, Exception):
|
||||
raw_stream_map[name] = res
|
||||
context_text = engine._format_stream_context(res)
|
||||
|
||||
# WP-25a: Automatisierte Kompression
|
||||
stream_cfg = library.get(name, {})
|
||||
threshold = stream_cfg.get("compression_threshold", 4000)
|
||||
|
||||
if len(context_text) > threshold:
|
||||
profile = stream_cfg.get("compression_profile")
|
||||
formatted_context_tasks.append(
|
||||
engine._compress_stream_content(name, context_text, request.message, profile)
|
||||
)
|
||||
else:
|
||||
# WP-20: Restaurierter Throttling-Schutz als Fallback
|
||||
if provider == "ollama" and len(context_text) > max_chars:
|
||||
context_text = context_text[:max_chars] + "\n[...]"
|
||||
|
||||
async def _ident(c=context_text): return c
|
||||
formatted_context_tasks.append(_ident())
|
||||
else:
|
||||
async def _err(): return "[Stream Error]"
|
||||
formatted_context_tasks.append(_err())
|
||||
|
||||
sources_hits = _collect_all_hits(raw_stream_map)
|
||||
|
||||
# Inhalte parallel finalisieren
|
||||
final_contexts = await asyncio.gather(*formatted_context_tasks)
|
||||
formatted_context_map = dict(zip(active_streams, final_contexts))
|
||||
|
||||
# Phase C: MoE Synthese
|
||||
# Phase C: Finale MoE Synthese
|
||||
answer_text = await engine._generate_final_answer(
|
||||
intent, strategy, request.message, formatted_context_map
|
||||
)
|
||||
sources_hits = _collect_all_hits(raw_stream_map)
|
||||
|
||||
duration_ms = int((time.time() - start_time) * 1000)
|
||||
|
||||
# Logging
|
||||
# Logging (WP-15)
|
||||
try:
|
||||
log_search(
|
||||
query_id=query_id, query_text=request.message, results=sources_hits,
|
||||
@@ -281,4 +254,4 @@ async def chat_endpoint(
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"❌ Chat Endpoint Failure: {e}", exc_info=True)
|
||||
raise HTTPException(status_code=500, detail="Fehler bei der Verarbeitung.")
|
||||
raise HTTPException(status_code=500, detail="Fehler bei der Verarbeitung der Anfrage.")
|
||||
Reference in new issue
Block a user