breakpilot-compliance/backend-compliance/compliance/services/checkers/embedding_checker.py

"""CONTENT-Pruefer / decision_method=EMBEDDING.

Ist die Pflicht SEMANTISCH im Text vorhanden? Max-Cosinus (Doc-Chunks x Control-
Paraphrasen) >= per-Control-Schwelle. Deterministisch (festes Embedding-Modell)
und gecacht. Rettet Recall-FP (Klausel da, anders formuliert).

Faellt der Embedding-Service aus, liefert der Checker present=None (unklar) — der
Aufrufer behaelt dann das Keyword-Ergebnis (kein Hang, kein Crash).
(Validiert an AGB: 17 Items, per-Item-Schwelle, 0 Fehl-Rescue.)
"""
from __future__ import annotations

import asyncio
import logging

from .base import CheckResult, ControlSpec, DocContext, VerificationMethod

logger = logging.getLogger(__name__)

# Paraphrasen-Vektoren je Control einmal einbetten + cachen.
_PARA_CACHE: dict[str, list] = {}


class EmbeddingChecker:
    verification_method = VerificationMethod.CONTENT

    async def check(self, ctrl: ControlSpec, doc: DocContext) -> CheckResult:
        text = doc.text or ""
        paras = ctrl.paraphrases or []
        thr = ctrl.embed_threshold if ctrl.embed_threshold is not None else 0.60
        if not paras or len(text) < 100:
            return CheckResult(present=None, source="embedding")
        try:
            from compliance.services.mc_embedding_matcher import (
                DIM, _chunk_text, _cosine, _embed_texts,
            )
            if ctrl.control_id not in _PARA_CACHE:
                pv = await _embed_texts(paras)
                _PARA_CACHE[ctrl.control_id] = [v for v in pv if v and len(v) == DIM]
            pvecs = _PARA_CACHE[ctrl.control_id]
            chunks = _chunk_text(text)
            cvecs = [v for v in await asyncio.wait_for(
                _embed_texts(chunks), timeout=90.0) if v and len(v) == DIM]
        except (Exception, asyncio.TimeoutError) as e:
            logger.info("embedding checker inaktiv %s: %s", ctrl.control_id, str(e)[:80])
            return CheckResult(present=None, source="embedding")
        if not pvecs or not cvecs:
            return CheckResult(present=None, source="embedding")
        best = max((_cosine(p, c) for p in pvecs for c in cvecs), default=0.0)
        return CheckResult(present=best >= thr, confidence=round(best, 3),
                           source="embedding")