mindnet/app/services/semantic_analyzer.py

"""
FILE: app/services/semantic_analyzer.py
DESCRIPTION: KI-gestützte Kanten-Validierung. Nutzt LLM (Background-Priority), um Kanten präzise einem Chunk zuzuordnen.
             WP-20 Fix: Volle Kompatibilität mit der provider-basierten Routing-Logik (OpenRouter Primary).
             WP-22: Integration von valid_types zur Halluzinations-Vermeidung.
VERSION: 2.2.3
STATUS: Active
DEPENDENCIES: app.services.llm_service, app.services.edge_registry, json, logging
LAST_ANALYSIS: 2025-12-24
"""

import json
import logging
from typing import List, Optional
from dataclasses import dataclass

# Importe
from app.services.llm_service import LLMService
# WP-22: Registry für Vokabular-Erzwingung
from app.services.edge_registry import registry as edge_registry

logger = logging.getLogger(__name__)

class SemanticAnalyzer:
    def __init__(self):
        self.llm = LLMService()

    def _is_valid_edge_string(self, edge_str: str) -> bool:
        """
        Prüft, ob ein String eine valide Kante im Format 'kind:target' ist.
        Verhindert, dass LLM-Geschwätz als Kante durchrutscht.
        """
        if not isinstance(edge_str, str) or ":" not in edge_str:
            return False
            
        parts = edge_str.split(":", 1)
        kind = parts[0].strip()
        target = parts[1].strip()
        
        # Regel 1: Ein 'kind' (Beziehungstyp) darf keine Leerzeichen enthalten.
        if " " in kind:
            return False
            
        # Regel 2: Plausible Länge für den Typ
        if len(kind) > 40 or len(kind) < 2:
            return False
            
        # Regel 3: Target darf nicht leer sein
        if not target:
            return False
            
        return True

    async def assign_edges_to_chunk(self, chunk_text: str, all_edges: List[str], note_type: str) -> List[str]:
        """
        Sendet einen Chunk und eine Liste potenzieller Kanten an das LLM.
        WP-20: Nutzt primär den Provider aus MINDNET_LLM_PROVIDER (OpenRouter).
        """
        if not all_edges:
            return []

        # 1. Bestimmung des Providers und Modells (WP-20)
        # Wir ziehen die Werte direkt aus dem Service-Kontext
        provider = self.llm.settings.MINDNET_LLM_PROVIDER
        model = self.llm.settings.OPENROUTER_MODEL if provider == "openrouter" else None

        # 2. Prompt laden via get_prompt
        prompt_template = self.llm.get_prompt("edge_allocation_template", provider)
        
        if not prompt_template or isinstance(prompt_template, dict):
            logger.warning("⚠️ [SemanticAnalyzer] Prompt 'edge_allocation_template' konnte nicht als String geladen werden. Nutze Not-Fallback.")
            prompt_template = (
                "TASK: Wähle aus den Kandidaten die relevanten Kanten für den Text.\n"
                "TEXT: {chunk_text}\n"
                "KANDIDATEN: {edge_list}\n"
                "OUTPUT: JSON Liste von Strings [\"kind:target\"]."
            )

        # 3. Daten für Template vorbereiten (WP-22 Integration)
        edge_registry.ensure_latest()
        valid_types_str = ", ".join(sorted(list(edge_registry.valid_types)))
        edges_str = "\n".join([f"- {e}" for e in all_edges])
        
        logger.debug(f"🔍 [SemanticAnalyzer] Request: {len(chunk_text)} chars Text, {len(all_edges)} Candidates.")

        # 4. Prompt füllen (FIX: valid_types hinzugefügt, um Format Error zu beheben)
        try:
            final_prompt = prompt_template.format(
                chunk_text=chunk_text[:3500], 
                edge_list=edges_str,
                valid_types=valid_types_str
            )
        except Exception as format_err:
            logger.error(f"❌ [SemanticAnalyzer] Format Error im Prompt-Template: {format_err}")
            return []

        try:
            # 5. LLM Call mit Traffic Control (Background Priority)
            response_json = await self.llm.generate_raw_response(
                prompt=final_prompt,
                force_json=True,
                max_retries=5, 
                base_delay=5.0,
                priority="background",
                provider=provider,
                model_override=model
            )

            logger.debug(f"📥 [SemanticAnalyzer] Raw Response (Preview): {response_json[:200]}...")

            # 6. Parsing & Cleaning
            clean_json = response_json.replace("```json", "").replace("```", "").strip()
            
            if not clean_json: 
                return []

            try:
                data = json.loads(clean_json)
            except json.JSONDecodeError as json_err:
                logger.error(f"❌ [SemanticAnalyzer] JSON Decode Error: {json_err}")
                return []

            valid_edges = []

            # 7. Robuste Validierung (List vs Dict)
            raw_candidates = []
            if isinstance(data, list):
                raw_candidates = data
            elif isinstance(data, dict):
                logger.info(f"ℹ️ [SemanticAnalyzer] LLM lieferte Dict statt Liste. Versuche Reparatur.")
                for key, val in data.items():
                    if key.lower() in ["edges", "results", "kanten", "matches"] and isinstance(val, list):
                         raw_candidates.extend(val)
                    elif isinstance(val, str):
                        raw_candidates.append(f"{key}:{val}")
                    elif isinstance(val, list):
                        for target in val:
                            if isinstance(target, str):
                                raw_candidates.append(f"{key}:{target}")

            # 8. Strict Validation Loop
            for e in raw_candidates:
                e_str = str(e)
                if self._is_valid_edge_string(e_str):
                    valid_edges.append(e_str)
                else:
                    logger.debug(f"   [SemanticAnalyzer] Invalid edge format rejected: '{e_str}'")

            final_result = [e for e in valid_edges if ":" in e]
            
            if final_result:
                logger.info(f"✅ [SemanticAnalyzer] Success. {len(final_result)} Kanten zugewiesen.")
            return final_result

        except Exception as e:
            logger.error(f"💥 [SemanticAnalyzer] Kritischer Fehler: {e}", exc_info=True)
            return []

    async def close(self):
        if self.llm:
            await self.llm.close()

# Singleton Helper
_analyzer_instance = None
def get_semantic_analyzer():
    global _analyzer_instance
    if _analyzer_instance is None:
        _analyzer_instance = SemanticAnalyzer()
    return _analyzer_instance