47 lines
1.4 KiB
Python
47 lines
1.4 KiB
Python
|
"""
|
|||
|
|
app/services/embeddings_client.py — Text→Embedding (WP-04)
|
|||
|
|
|
|||
|
|
Zweck:
|
|||
|
|
Liefert 384-d Embeddings für Textqueries (lazy load, einmal pro Prozess).
|
|||
|
|
Standard: Sentence-Transformers (MODEL_NAME aus app.config.Settings).
|
|||
|
|
Hinweis: Kein Netz-Zugriff; nutzt lokal installierte Modelle.
|
|||
|
|
Kompatibilität:
|
|||
|
|
Python 3.12+, sentence-transformers 5.x
|
|||
|
|
Version:
|
|||
|
|
0.1.0 (Erstanlage)
|
|||
|
|
Stand:
|
|||
|
|
2025-10-07
|
|||
|
|
Bezug:
|
|||
|
|
- app/core/retriever.py (nutzt embed_text_if_needed)
|
|||
|
|
- app/config.py (MODEL_NAME, VECTOR_SIZE)
|
|||
|
|
Nutzung:
|
|||
|
|
from app.services.embeddings_client import embed_text
|
|||
|
|
Änderungsverlauf:
|
|||
|
|
0.1.0 (2025-10-07) – Erstanlage.
|
|||
|
|
"""
|
|||
|
|
|
|||
|
|
from __future__ import annotations
|
|||
|
|
from typing import List
|
|||
|
|
from functools import lru_cache
|
|||
|
|
from app.config import get_settings
|
|||
|
|
|
|||
|
|
# Lazy import, damit Testläufe ohne Modell-Laden schnell sind
|
|||
|
|
def _load_model():
|
|||
|
|
from sentence_transformers import SentenceTransformer # import hier, nicht top-level
|
|||
|
|
s = get_settings()
|
|||
|
|
return SentenceTransformer(s.MODEL_NAME, device="cpu")
|
|||
|
|
|
|||
|
|
@lru_cache(maxsize=1)
|
|||
|
|
def _cached_model():
|
|||
|
|
return _load_model()
|
|||
|
|
|
|||
|
|
def embed_text(text: str) -> List[float]:
|
|||
|
|
"""
|
|||
|
|
Erzeugt einen 384-d Vektor (oder laut Settings.VECTOR_SIZE) für den gegebenen Text.
|
|||
|
|
"""
|
|||
|
|
if not text or not text.strip():
|
|||
|
|
raise ValueError("embed_text: leerer Text")
|
|||
|
|
model = _cached_model()
|
|||
|
|
vec = model.encode([text], normalize_embeddings=True)[0]
|
|||
|
|
return vec.astype(float).tolist()
|