Integrate symmetric edge logic and discovery API: Update ingestion processor and validation to support automatic inverse edge generation. Enhance edge registry for dual vocabulary and schema management. Introduce new discovery endpoint for proactive edge suggestions, improving graph topology and edge validation processes.
This commit is contained in:
1 parent
745352ff3f
commit
4802eba27b
4 files changed
+344
-109
No files matched your search
+109
-7
@@ -3,9 +3,11 @@ FILE: app/routers/chat.py
|
||||
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)
|
||||
WP-24c: Integration der Discovery API für proaktive Vernetzung.
|
||||
VERSION: 3.1.0 (WP-24c: Discovery API Integration)
|
||||
STATUS: Active
|
||||
FIX:
|
||||
- WP-24c: Neuer Endpunkt /query/discover für proaktive Kanten-Vorschläge.
|
||||
- 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).
|
||||
@@ -13,6 +15,7 @@ FIX:
|
||||
|
||||
from fastapi import APIRouter, HTTPException, Depends
|
||||
from typing import List, Dict, Any, Optional
|
||||
from pydantic import BaseModel
|
||||
import time
|
||||
import uuid
|
||||
import logging
|
||||
@@ -22,13 +25,27 @@ import asyncio
|
||||
from pathlib import Path
|
||||
|
||||
from app.config import get_settings
|
||||
from app.models.dto import ChatRequest, ChatResponse, QueryHit
|
||||
from app.models.dto import ChatRequest, ChatResponse, QueryHit, QueryRequest
|
||||
from app.services.llm_service import LLMService
|
||||
from app.services.feedback_service import log_search
|
||||
|
||||
router = APIRouter()
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
# --- EBENE 0: DTOs FÜR DISCOVERY (WP-24c) ---
|
||||
|
||||
class DiscoveryRequest(BaseModel):
|
||||
content: str
|
||||
top_k: int = 8
|
||||
min_confidence: float = 0.6
|
||||
|
||||
class DiscoveryHit(BaseModel):
|
||||
target_note: str # Note ID
|
||||
target_title: str # Menschenlesbarer Titel
|
||||
suggested_edge_type: str # Kanonischer Typ aus edge_vocabulary
|
||||
confidence_score: float # Kombinierter Vektor- + KI-Score
|
||||
reasoning: str # Kurze Begründung der KI
|
||||
|
||||
# --- EBENE 1: CONFIG LOADER & CACHING (WP-25 Standard) ---
|
||||
|
||||
_DECISION_CONFIG_CACHE = None
|
||||
@@ -135,8 +152,7 @@ async def _classify_intent(query: str, llm: LLMService) -> tuple[str, str]:
|
||||
return "INTERVIEW", "Keyword (Interview)"
|
||||
|
||||
# 3. SLOW PATH: DecisionEngine LLM Router (MoE-gesteuert)
|
||||
# WP-25b FIX: Nutzung der öffentlichen API statt privater Methode
|
||||
intent = await llm.decision_engine._determine_strategy(query) # TODO: Public API erstellen
|
||||
intent = await llm.decision_engine._determine_strategy(query)
|
||||
return intent, "DecisionEngine (LLM)"
|
||||
|
||||
# --- EBENE 3: RETRIEVAL AGGREGATION ---
|
||||
@@ -154,7 +170,7 @@ def _collect_all_hits(stream_responses: Dict[str, Any]) -> List[QueryHit]:
|
||||
seen_node_ids.add(hit.node_id)
|
||||
return sorted(all_hits, key=lambda h: h.total_score, reverse=True)
|
||||
|
||||
# --- EBENE 4: ENDPUNKT ---
|
||||
# --- EBENE 4: ENDPUNKTE ---
|
||||
|
||||
def get_llm_service():
|
||||
return LLMService()
|
||||
@@ -196,7 +212,6 @@ async def chat_endpoint(
|
||||
template_key = strategy.get("prompt_template", "interview_template")
|
||||
|
||||
# WP-25b: Lazy Loading Call
|
||||
# Wir übergeben nur Key und Variablen. Das System formatiert passend zum Modell.
|
||||
answer_text = await llm.generate_raw_response(
|
||||
prompt_key=template_key,
|
||||
variables={
|
||||
@@ -257,4 +272,91 @@ 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 der Anfrage.")
|
||||
raise HTTPException(status_code=500, detail="Fehler bei der Verarbeitung der Anfrage.")
|
||||
|
||||
@router.post("/query/discover", response_model=List[DiscoveryHit])
|
||||
async def discover_edges(
|
||||
request: DiscoveryRequest,
|
||||
llm: LLMService = Depends(get_llm_service)
|
||||
):
|
||||
"""
|
||||
WP-24c: Analysiert Text auf potenzielle Kanten zu bestehendem Wissen.
|
||||
Nutzt Vektor-Suche und DecisionEngine-Logik (WP-25b PROMPT-TRACE konform).
|
||||
"""
|
||||
start_time = time.time()
|
||||
logger.info(f"🔍 [WP-24c] Discovery triggered for content: {request.content[:50]}...")
|
||||
|
||||
try:
|
||||
# 1. Kandidaten-Suche via Retriever (Vektor-Match)
|
||||
search_req = QueryRequest(
|
||||
query=request.content,
|
||||
top_k=request.top_k,
|
||||
explain=True
|
||||
)
|
||||
candidates = await llm.decision_engine.retriever.search(search_req)
|
||||
|
||||
if not candidates.results:
|
||||
logger.info("ℹ️ No candidates found for discovery.")
|
||||
return []
|
||||
|
||||
# 2. KI-gestützte Beziehungs-Extraktion (WP-25b)
|
||||
discovery_results = []
|
||||
|
||||
# Zugriff auf gültige Kanten-Typen aus der Registry
|
||||
from app.services.edge_registry import registry as edge_reg
|
||||
valid_types_str = ", ".join(list(edge_reg.valid_types))
|
||||
|
||||
# Parallele Evaluierung der Kandidaten für maximale Performance
|
||||
async def evaluate_candidate(hit: QueryHit) -> Optional[DiscoveryHit]:
|
||||
if hit.total_score < request.min_confidence:
|
||||
return None
|
||||
|
||||
try:
|
||||
# Nutzt ingest_extractor Profil für präzise semantische Analyse
|
||||
# Wir verwenden das prompt_key Pattern (edge_extraction) gemäß WP-24c Vorgabe
|
||||
raw_suggestion = await llm.generate_raw_response(
|
||||
prompt_key="edge_extraction",
|
||||
variables={
|
||||
"note_id": "NEUER_INHALT",
|
||||
"text": f"PROXIMITY_TARGET: {hit.source.get('text', '')}\n\nNEW_CONTENT: {request.content}",
|
||||
"valid_types": valid_types_str
|
||||
},
|
||||
profile_name="ingest_extractor",
|
||||
priority="realtime"
|
||||
)
|
||||
|
||||
# Parsing der LLM Antwort (Erwartet JSON Liste)
|
||||
from app.core.ingestion.ingestion_utils import extract_json_from_response
|
||||
suggestions = extract_json_from_response(raw_suggestion)
|
||||
|
||||
if isinstance(suggestions, list) and len(suggestions) > 0:
|
||||
sugg = suggestions[0] # Wir nehmen den stärksten Vorschlag pro Hit
|
||||
return DiscoveryHit(
|
||||
target_note=hit.note_id,
|
||||
target_title=hit.source.get("title") or hit.note_id,
|
||||
suggested_edge_type=sugg.get("kind", "related_to"),
|
||||
confidence_score=hit.total_score,
|
||||
reasoning=f"Semantische Nähe ({int(hit.total_score*100)}%) entdeckt."
|
||||
)
|
||||
except Exception as e:
|
||||
logger.warning(f"⚠️ Discovery evaluation failed for hit {hit.note_id}: {e}")
|
||||
return None
|
||||
|
||||
tasks = [evaluate_candidate(hit) for hit in candidates.results]
|
||||
results = await asyncio.gather(*tasks)
|
||||
|
||||
# Zusammenführung und Duplikat-Bereinigung
|
||||
seen_targets = set()
|
||||
for r in results:
|
||||
if r and r.target_note not in seen_targets:
|
||||
discovery_results.append(r)
|
||||
seen_targets.add(r.target_note)
|
||||
|
||||
duration = int((time.time() - start_time) * 1000)
|
||||
logger.info(f"✨ Discovery finished: found {len(discovery_results)} edges in {duration}ms")
|
||||
|
||||
return discovery_results
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"❌ Discovery API failure: {e}", exc_info=True)
|
||||
raise HTTPException(status_code=500, detail="Discovery-Prozess fehlgeschlagen.")
|
||||
Reference in new issue
Block a user