WP24c - Agentic Edge Validation & Chunk-Aware Multigraph-System (v4.5.8) #22

Merged
Lars merged 71 commits from WP24c into main 2026-01-12 10:53:20 +01:00
Showing only changes of commit 72cf71fa87 - Show all commits

View File

@ -5,14 +5,15 @@ DESCRIPTION: Der zentrale IngestionService (Orchestrator).
WP-25a: Integration der Mixture of Experts (MoE) Architektur.
WP-15b: Two-Pass Workflow mit globalem Kontext-Cache.
WP-20/22: Cloud-Resilienz und Content-Lifecycle integriert.
AUDIT v3.1.7: Explicit Authority Enforcement. Verhindert durch interne
ID-Registry und DB-Abgleich das Überschreiben manueller Kanten.
VERSION: 3.1.7 (WP-24c: Strict Authority Protection)
AUDIT v3.1.8: Fix für HTTP 400 (Bad Request) durch ID-Validierung
und Schutz vor System-Typ Kollisionen.
VERSION: 3.1.8 (WP-24c: Robust Symmetry & ID Validation)
STATUS: Active
"""
import logging
import asyncio
import os
import re
from typing import Dict, List, Optional, Tuple, Any
# Core Module Imports
@ -27,7 +28,7 @@ from app.core.graph.graph_utils import _mk_edge_id
# MODULARISIERUNG: Neue Import-Pfade für die Datenbank-Ebene
from app.core.database.qdrant import QdrantConfig, get_client, ensure_collections, ensure_payload_indexes
from app.core.database.qdrant_points import points_for_chunks, points_for_note, points_for_edges, upsert_batch
from qdrant_client.http import models as rest # Für Real-Time DB-Checks
from qdrant_client.http import models as rest
# Services
from app.services.embeddings_client import EmbeddingsClient
@ -36,7 +37,7 @@ from app.services.llm_service import LLMService
# Package-Interne Imports
from .ingestion_utils import load_type_registry, resolve_note_type, get_chunk_config_by_profile
from .ingestion_db import fetch_note_payload, artifacts_missing, purge_artifacts
from .ingestion_db import fetch_note_payload, artifacts_missing, purge_artifacts, is_explicit_edge_present
from .ingestion_validation import validate_edge_candidate
from .ingestion_note_payload import make_note_payload
from .ingestion_chunk_payload import make_chunk_payloads
@ -83,10 +84,35 @@ class IngestionService:
except Exception as e:
logger.warning(f"DB initialization warning: {e}")
def _is_valid_note_id(self, text: str) -> bool:
"""
Prüft, ob ein String eine plausible Note-ID oder ein gültiger Titel ist.
Verhindert Symmetrie-Kanten zu Typ-Strings wie 'insight', 'event' oder 'source'.
"""
if not text or len(text.strip()) < 3:
return False
# 1. Bekannte System-Typen oder Meta-Daten Begriffe ausschließen
# Diese landen oft durch fehlerhafte Frontmatter-Einträge in der Referenz-Liste
blacklisted = {
"insight", "event", "source", "task", "project",
"person", "concept", "value", "principle", "trip",
"lesson", "decision", "requirement", "related_to"
}
clean_text = text.lower().strip()
if clean_text in blacklisted:
return False
# 2. Ausschluss von zu langen Textfragmenten (wahrscheinlich kein Titel/ID)
if len(text) > 120:
return False
return True
async def run_batch(self, file_paths: List[str], vault_root: str) -> List[Dict[str, Any]]:
"""
WP-15b: Implementiert den Two-Pass Ingestion Workflow.
Pass 1: Pre-Scan füllt den Context-Cache (3-Wege-Indexierung).
Pass 1: Pre-Scan füllt den Context-Cache.
Pass 2: Verarbeitung nutzt den Cache für die semantische Prüfung.
"""
# Reset der Authority-Registry für den neuen Batch
@ -109,26 +135,6 @@ class IngestionService:
logger.info(f"🚀 [Pass 2] Semantic Processing of {len(file_paths)} files...")
return [await self.process_file(p, vault_root, apply=True, purge_before=True) for p in file_paths]
async def _is_explicit_edge_in_db(self, edge_id: str) -> bool:
"""
WP-24c: Prüft via Point-ID, ob bereits eine explizite (manuelle) Kante in Qdrant liegt.
Verhindert, dass virtuelle Symmetrien bestehendes Wissen überschreiben.
"""
edges_col = f"{self.prefix}_edges"
try:
# Direkte Punkt-Abfrage ist schneller als Scroll/Filter
res = self.client.retrieve(
collection_name=edges_col,
ids=[edge_id],
with_payload=True,
with_vectors=False
)
if res and not res[0].payload.get("virtual", False):
return True # Punkt existiert und ist NICHT virtuell
return False
except Exception:
return False
async def process_file(self, file_path: str, vault_root: str, **kwargs) -> Dict[str, Any]:
"""Transformiert eine Markdown-Datei in den Graphen."""
apply = kwargs.get("apply", False)
@ -149,7 +155,7 @@ class IngestionService:
except Exception as e:
return {**result, "error": f"Validation failed: {str(e)}"}
# Dynamischer Lifecycle-Filter aus der Registry (WP-14)
# Dynamischer Lifecycle-Filter
ingest_cfg = self.registry.get("ingestion_settings", {})
ignore_list = ingest_cfg.get("ignore_statuses", ["system", "template", "archive", "hidden"])
@ -157,7 +163,7 @@ class IngestionService:
if current_status in ignore_list:
return {**result, "status": "skipped", "reason": "lifecycle_filter"}
# 2. Payload & Change Detection (Multi-Hash)
# 2. Payload & Change Detection
note_type = resolve_note_type(self.registry, fm.get("type"))
note_pl = make_note_payload(
parsed, vault_root=vault_root, file_path=file_path,
@ -166,43 +172,37 @@ class IngestionService:
)
note_id = note_pl["note_id"]
# Abgleich mit der Datenbank (Qdrant)
# Abgleich mit der Datenbank
old_payload = None if force_replace else fetch_note_payload(self.client, self.prefix, note_id)
# Prüfung gegen den konfigurierten Hash-Modus (body vs. full)
# Prüfung gegen den konfigurierten Hash-Modus
check_key = f"{self.active_hash_mode}:{hash_source}:{hash_normalize}"
old_hash = (old_payload or {}).get("hashes", {}).get(check_key)
new_hash = note_pl.get("hashes", {}).get(check_key)
# Check ob Chunks oder Kanten in der DB fehlen (Reparatur-Modus)
c_miss, e_miss = artifacts_missing(self.client, self.prefix, note_id)
# Wenn Hash identisch und Artefakte vorhanden -> Skip
if not (force_replace or not old_payload or old_hash != new_hash or c_miss or e_miss):
return {**result, "status": "unchanged", "note_id": note_id}
if not apply:
return {**result, "status": "dry-run", "changed": True, "note_id": note_id}
# 3. Deep Processing (Chunking, Validation, Embedding)
# 3. Deep Processing
try:
body_text = getattr(parsed, "body", "") or ""
edge_registry.ensure_latest()
# Profil-Auflösung via Registry
profile = note_pl.get("chunk_profile", "sliding_standard")
chunk_cfg = get_chunk_config_by_profile(self.registry, profile, note_type)
enable_smart = chunk_cfg.get("enable_smart_edge_allocation", False)
# WP-15b: Chunker-Aufruf bereitet den Candidate-Pool pro Chunk vor.
chunks = await assemble_chunks(note_id, body_text, note_type, config=chunk_cfg)
# Semantische Kanten-Validierung (Primärprüfung)
# Semantische Kanten-Validierung
for ch in chunks:
new_pool = []
for cand in getattr(ch, "candidate_pool", []):
# WP-25a: Profilgesteuerte binäre Validierung
if cand.get("provenance") == "global_pool" and enable_smart:
is_valid = await validate_edge_candidate(
ch.text,
@ -214,20 +214,17 @@ class IngestionService:
if is_valid:
new_pool.append(cand)
else:
# Explizite Kanten (Wikilinks/Callouts) werden übernommen
new_pool.append(cand)
ch.candidate_pool = new_pool
# Payload-Erstellung für die Chunks
chunk_pls = make_chunk_payloads(
fm, note_pl["path"], chunks, file_path=file_path,
types_cfg=self.registry
)
# Vektorisierung der Fenster-Texte
vecs = await self.embedder.embed_documents([c.get("window") or "" for c in chunk_pls]) if chunk_pls else []
# Aggregation aller finalen Kanten (Edges)
# Aggregation aller Kanten
raw_edges = build_edges_for_note(
note_id, chunk_pls,
note_level_references=note_pl.get("references", []),
@ -239,25 +236,29 @@ class IngestionService:
# PHASE 1: Alle expliziten Kanten vorverarbeiten und registrieren
for e in raw_edges:
target_raw = e.get("target_id")
# Robustheits-Check: Ist das Ziel eine valide Note-ID?
if not self._is_valid_note_id(target_raw):
continue
resolved_kind = edge_registry.resolve(
e.get("kind", "related_to"),
provenance=e.get("provenance", "explicit"),
context={"file": file_path, "note_id": note_id}
)
e["kind"] = resolved_kind
# Markierung der Herkunft für selektiven Purge
e["origin_note_id"] = note_id
e["virtual"] = False # Authority-Markierung für explizite Kanten
e["virtual"] = False
e["confidence"] = e.get("confidence", 1.0) # Volle Gewichtung
# Registrierung der ID im Laufzeit-Schutz (Authority)
edge_id = _mk_edge_id(resolved_kind, note_id, e.get("target_id"), e.get("scope", "note"))
edge_id = _mk_edge_id(resolved_kind, note_id, target_raw, e.get("scope", "note"))
self.processed_explicit_ids.add(edge_id)
final_edges.append(e)
# PHASE 2: Symmetrische Kanten (Invers) mit Authority-Schutz erzeugen
# Wir nutzen hierfür nur die expliziten Kanten aus Phase 1 als Basis
explicit_only = [x for x in final_edges if not x.get("virtual")]
for e in explicit_only:
@ -265,63 +266,53 @@ class IngestionService:
inverse_kind = edge_registry.get_inverse(kind)
target_raw = e.get("target_id")
# ID-Resolution: Finden der echten Note_ID im Cache
# ID-Resolution
target_ctx = self.batch_cache.get(target_raw)
target_canonical_id = target_ctx.note_id if target_ctx else target_raw
# Validierung für Symmetrie-Erzeugung (Kein Self-Loop, Existenz der Inversen)
if (inverse_kind and target_canonical_id and target_canonical_id != note_id):
# Validierung für Symmetrie-Erzeugung (Kein Self-Loop, valide ID)
if (inverse_kind and target_canonical_id and target_canonical_id != note_id and self._is_valid_note_id(target_canonical_id)):
# 1. ID der potenziellen virtuellen Kante berechnen
# Wir nutzen exakt die Parameter, die auch points_for_edges nutzt
potential_id = _mk_edge_id(inverse_kind, target_canonical_id, note_id, e.get("scope", "note"))
# 2. AUTHORITY-CHECK A: Wurde diese Kante bereits explizit im aktuellen Batch registriert?
# AUTHORITY-CHECK: Batch-Gedächtnis oder Datenbank
is_in_batch = potential_id in self.processed_explicit_ids
# 3. AUTHORITY-CHECK B: Existiert sie bereits als explizit in der Datenbank?
is_in_db = False
if not is_in_batch:
is_in_db = await self._is_explicit_edge_in_db(potential_id)
# Real-Time DB Check verhindert 400 Bad Request durch vorherige ID-Validierung
is_in_db = await is_explicit_edge_present(self.client, self.prefix, potential_id)
# 4. Filter: Nur anlegen, wenn KEINE explizite Autorität vorliegt
# Keine Abwertung der Confidence auf Wunsch des Nutzers
if not is_in_batch and not is_in_db:
if (inverse_kind != kind or kind not in ["related_to", "references"]):
inv_edge = e.copy()
# Richtungs-Umkehr
inv_edge["note_id"] = target_canonical_id
inv_edge["target_id"] = note_id
inv_edge["kind"] = inverse_kind
# Metadaten für Struktur-Kante
inv_edge["virtual"] = True
inv_edge["provenance"] = "structure"
inv_edge["confidence"] = e.get("confidence", 1.0) # Gewichtung bleibt gleich
# Lifecycle-Verankerung: Diese Kante gehört logisch zum Verursacher (Note A)
inv_edge["origin_note_id"] = note_id
inv_edge.update({
"note_id": target_canonical_id,
"target_id": note_id,
"kind": inverse_kind,
"virtual": True,
"provenance": "structure",
"confidence": 1.0, # Gewichtung bleibt gleich laut Nutzerwunsch
"origin_note_id": note_id
})
final_edges.append(inv_edge)
logger.info(f"🔄 [SYMMETRY] Built inverse: {target_canonical_id} --({inverse_kind})--> {note_id}")
edges = final_edges
# 4. DB Upsert via modularisierter Points-Logik
# 4. DB Upsert
if apply:
if purge_before and old_payload:
purge_artifacts(self.client, self.prefix, note_id)
# Speichern der Haupt-Note
# Speichern
n_name, n_pts = points_for_note(self.prefix, note_pl, None, self.dim)
upsert_batch(self.client, n_name, n_pts)
# Speichern der Chunks
if chunk_pls and vecs:
c_pts = points_for_chunks(self.prefix, chunk_pls, vecs)[1]
upsert_batch(self.client, f"{self.prefix}_chunks", c_pts)
# Speichern der Kanten (inklusive der virtuellen Inversen)
if edges:
e_pts = points_for_edges(self.prefix, edges)[1]
upsert_batch(self.client, f"{self.prefix}_edges", e_pts)
@ -339,11 +330,10 @@ class IngestionService:
return {**result, "error": str(e)}
async def create_from_text(self, markdown_content: str, filename: str, vault_root: str, folder: str = "00_Inbox") -> Dict[str, Any]:
"""Erstellt eine Note aus einem Textstream und triggert die Ingestion."""
"""Erstellt eine Note aus einem Textstream."""
target_path = os.path.join(vault_root, folder, filename)
os.makedirs(os.path.dirname(target_path), exist_ok=True)
with open(target_path, "w", encoding="utf-8") as f:
f.write(markdown_content)
await asyncio.sleep(0.1)
# Triggert sofortigen Import mit force_replace/purge_before
return await self.process_file(file_path=target_path, vault_root=vault_root, apply=True, force_replace=True, purge_before=True)