112 lines
4.0 KiB
Python
112 lines
4.0 KiB
Python
|
"""QwenBewerter — Produktions-Adapter für den LlmBewerter-Port.
|
||
|
|
|
||
|
|
Kapselt den ``AsyncOpenAI``-Client gegen die DashScope-API, den Retry-
|
||
|
|
Loop mit Temperatur-Escalation und das Markdown-Fence-Stripping. Die
|
||
|
|
Retry-Semantik bleibt identisch zu ``analyzer.py`` vor der Migration:
|
||
|
|
bis zu ``max_retries`` Versuche, Temperatur steigt um 0.1 pro Versuch.
|
||
|
|
|
||
|
|
Der Adapter gibt den geparsten ``dict`` zurück — Pydantic-Validierung,
|
||
|
|
Citation-Binding und Missing-Programme-Check bleiben Sache des Callers
|
||
|
|
in ``analyzer.py``.
|
||
|
|
"""
|
||
|
|
|
||
|
|
from __future__ import annotations
|
||
|
|
|
||
|
|
import hashlib
|
||
|
|
import json
|
||
|
|
import logging
|
||
|
|
from typing import Optional
|
||
|
|
|
||
|
|
from ..config import settings
|
||
|
|
from ..ports.llm_bewerter import LlmRequest
|
||
|
|
|
||
|
|
logger = logging.getLogger(__name__)
|
||
|
|
|
||
|
|
|
||
|
|
def _content_fingerprint(content: str) -> str:
|
||
|
|
"""Log-sicherer Identifier ohne PII-Leak (Issue #57 Befund #4)."""
|
||
|
|
if not content:
|
||
|
|
return "len=0"
|
||
|
|
h = hashlib.sha1(content.encode("utf-8", errors="replace")).hexdigest()[:8]
|
||
|
|
return f"len={len(content)} sha1={h}"
|
||
|
|
|
||
|
|
|
||
|
|
def _strip_markdown_fences(content: str) -> str:
|
||
|
|
"""Entfernt Markdown-Code-Fences, die Qwen trotz Prompt manchmal ergänzt.
|
||
|
|
|
||
|
|
In Sync mit ``analyzer.py`` vor der Migration; Einheitstests in
|
||
|
|
``tests/test_analyzer.py`` spiegeln exakt diese Logik.
|
||
|
|
"""
|
||
|
|
content = content.strip()
|
||
|
|
if content.startswith("```"):
|
||
|
|
content = content.split("\n", 1)[1]
|
||
|
|
if content.endswith("```"):
|
||
|
|
content = content.rsplit("```", 1)[0]
|
||
|
|
if content.startswith("```json"):
|
||
|
|
content = content[7:]
|
||
|
|
return content.strip()
|
||
|
|
|
||
|
|
|
||
|
|
class QwenBewerter:
|
||
|
|
"""LlmBewerter-Adapter für Qwen Plus (via DashScope)."""
|
||
|
|
|
||
|
|
def __init__(
|
||
|
|
self,
|
||
|
|
api_key: Optional[str] = None,
|
||
|
|
base_url: Optional[str] = None,
|
||
|
|
client=None,
|
||
|
|
) -> None:
|
||
|
|
"""Konstruktor-Injection erlaubt Tests, einen Mock-Client zu reichen
|
||
|
|
ohne Netzwerk-Zugriff. Prod nutzt den Default: Settings aus
|
||
|
|
``config.py`` + ``AsyncOpenAI``."""
|
||
|
|
self._api_key = api_key or settings.dashscope_api_key
|
||
|
|
self._base_url = base_url or settings.dashscope_base_url
|
||
|
|
self._client = client # lazy-created in .bewerte() wenn nicht gesetzt
|
||
|
|
|
||
|
|
def _get_client(self):
|
||
|
|
if self._client is not None:
|
||
|
|
return self._client
|
||
|
|
# Lazy-Import, damit die Test-Suite ohne ``openai``-Paket laufen kann.
|
||
|
|
from openai import AsyncOpenAI
|
||
|
|
|
||
|
|
self._client = AsyncOpenAI(api_key=self._api_key, base_url=self._base_url)
|
||
|
|
return self._client
|
||
|
|
|
||
|
|
async def bewerte(self, request: LlmRequest) -> dict:
|
||
|
|
"""Führt den LLM-Call aus, bis JSON-Parse klappt oder Retries erschöpft."""
|
||
|
|
client = self._get_client()
|
||
|
|
|
||
|
|
last_error: Optional[Exception] = None
|
||
|
|
for attempt in range(request.max_retries):
|
||
|
|
response = await client.chat.completions.create(
|
||
|
|
model=request.model,
|
||
|
|
messages=[
|
||
|
|
{"role": "system", "content": request.system_prompt},
|
||
|
|
{"role": "user", "content": request.user_prompt},
|
||
|
|
],
|
||
|
|
temperature=request.base_temperature + (attempt * 0.1),
|
||
|
|
max_tokens=request.max_tokens,
|
||
|
|
)
|
||
|
|
content = response.choices[0].message.content.strip()
|
||
|
|
content = _strip_markdown_fences(content)
|
||
|
|
|
||
|
|
try:
|
||
|
|
return json.loads(content)
|
||
|
|
except json.JSONDecodeError as e:
|
||
|
|
last_error = e
|
||
|
|
logger.warning(
|
||
|
|
"LLM JSON parse error attempt %d/%d (%s) — content %s",
|
||
|
|
attempt + 1, request.max_retries, e,
|
||
|
|
_content_fingerprint(content),
|
||
|
|
)
|
||
|
|
if attempt >= request.max_retries - 1:
|
||
|
|
logger.error(
|
||
|
|
"LLM JSON parsing exhausted retries, content %s",
|
||
|
|
_content_fingerprint(content),
|
||
|
|
)
|
||
|
|
raise
|
||
|
|
|
||
|
|
# Unreachable — letzter Versuch hat raised. Für Typcheck.
|
||
|
|
assert last_error is not None
|
||
|
|
raise last_error
|