mindnet/config/types.yaml

version: 1.2 # Update für Smart Chunking Config

# --- CHUNKING DEFINITIONEN ---
# Hier definieren wir die technischen Strategien zentral.
chunking_profiles:
  # Standard für Fließtexte (Sliding Window)
  sliding_short:
    strategy: sliding_window
    target: 200
    max: 350
    overlap: [30, 50]
  
  sliding_standard:
    strategy: sliding_window
    target: 400
    max: 600
    overlap: [50, 80]

  sliding_large:
    strategy: sliding_window
    target: 500
    max: 800
    overlap: [60, 100]

  # Smart Chunking für Strukturen (Harte Splits)
  structured_strict:
    strategy: by_heading
    split_level: 2
    max: 600              # Fallback Limit
    target: 400           # Fallback Target bei Sub-Chunking
    overlap: [50, 80]     # Overlap bei Sub-Chunking
    
  # NEU: LLM-basierte semantische Zerlegung (Chunker.py ruft semantic_analyzer.py)
  semantic_llm:
    strategy: semantic_llm
    # Da das LLM die Längensteuerung übernimmt, dienen diese als Fallback/Empfehlung
    target: 400 
    max: 800 
    
defaults:
  retriever_weight: 1.0
  chunking_profile: sliding_standard # Fallback Profil
  edge_defaults: [] 

types:
  # --- WISSENSBAUSTEINE ---
  concept:
    chunking_profile: sliding_standard
    retriever_weight: 0.60
    edge_defaults: ["references", "related_to"]

  source:
    chunking_profile: sliding_standard
    retriever_weight: 0.50
    edge_defaults: [] 

  glossary:
    chunking_profile: sliding_short
    retriever_weight: 0.40
    edge_defaults: ["related_to"]

  # --- IDENTITÄT & PERSÖNLICHKEIT ---
  profile:
    chunking_profile: structured_strict 
    retriever_weight: 0.70
    edge_defaults: ["references", "related_to"]

  value:
    chunking_profile: structured_strict
    retriever_weight: 1.00
    edge_defaults: ["related_to"]

  principle:
    chunking_profile: structured_strict
    retriever_weight: 0.95
    edge_defaults: ["derived_from", "references"]

  belief:
    chunking_profile: sliding_short
    retriever_weight: 0.90
    edge_defaults: ["related_to"]

  experience:
    chunking_profile: sliding_standard
    retriever_weight: 0.90
    edge_defaults: ["derived_from", "references"]

  # --- STRATEGIE & ENTSCHEIDUNG ---
  goal:
    chunking_profile: sliding_standard
    retriever_weight: 0.95
    edge_defaults: ["depends_on", "related_to"]

  decision:
    chunking_profile: structured_strict 
    retriever_weight: 1.00
    edge_defaults: ["caused_by", "references"]

  risk:
    chunking_profile: sliding_short
    retriever_weight: 0.85
    edge_defaults: ["related_to", "blocks"]

  milestone:
    chunking_profile: sliding_short
    retriever_weight: 0.70
    edge_defaults: ["related_to", "part_of"]

  # --- OPERATIV ---
  project:
    chunking_profile: sliding_large 
    retriever_weight: 0.97
    edge_defaults: ["references", "depends_on"] 

  task:
    chunking_profile: sliding_short
    retriever_weight: 0.80
    edge_defaults: ["depends_on", "part_of"]

  journal:
    # NEUE ZUWEISUNG: Journale profitieren am meisten von der semantischen Analyse
    chunking_profile: semantic_llm 
    retriever_weight: 0.80
    edge_defaults: ["references", "related_to"]