Onyscribe-Admin/app/services/llm.py

"""Zugang zu Cortecs (OpenAI-kompatibel + Anthropic-Messages) für Agent und Prompt-Test.

Der Key kommt aus dem geteilten Secret CORTECS_API_KEY (verschlüsselt in der DB) oder der Env CORTECS_API_KEY.
"""
from __future__ import annotations

import os

import httpx
from sqlalchemy import select
from sqlalchemy.orm import Session

from app.models import Secret, Setting
from app.services import crypto

DEFAULT_BASE = "https://api.cortecs.ai/v1"


class LLMNotConfigured(RuntimeError):
    pass


def _setting(db: Session, key: str, default: str = "") -> str:
    s = db.get(Setting, key)
    return s.value if s and s.value else default


def key_info(db: Session, tenant_id: str | None = None) -> tuple[str, str, str | None]:
    """(key, quelle, fehler). Reihenfolge: geteiltes Secret (DB) → Env CORTECS_API_KEY → Key des Tenants (Rückfall).
    Quelle: 'shared' | 'env' | 'tenant' | 'none'. fehler: z. B. nicht entschlüsselbar (SECRET_KEY geändert?)."""
    error = None
    rows = [("shared", db.scalar(select(Secret).where(Secret.tenant_id.is_(None), Secret.key == "CORTECS_API_KEY")))]
    for source, row in rows:
        if row and row.value_enc:
            try:
                return crypto.decrypt(row.value_enc), source, None
            except Exception:  # noqa: BLE001 – falscher SECRET_KEY o. Ä.
                error = "Geteilter CORTECS_API_KEY in der DB ist nicht entschlüsselbar (SECRET_KEY geändert?) – bitte neu setzen."
    env = os.getenv("CORTECS_API_KEY", "").strip()
    if env:
        return env, "env", error
    if tenant_id:
        row = db.scalar(select(Secret).where(Secret.tenant_id == tenant_id, Secret.key == "CORTECS_API_KEY"))
        if row and row.value_enc:
            try:
                return crypto.decrypt(row.value_enc), "tenant", error
            except Exception:  # noqa: BLE001
                error = error or "CORTECS_API_KEY dieses Tenants ist nicht entschlüsselbar."
    return "", "none", error


def api_key(db: Session, tenant_id: str | None = None) -> str:
    return key_info(db, tenant_id)[0]


def base_url(db: Session) -> str:
    return _setting(db, "cortecs_base_url", DEFAULT_BASE).rstrip("/")


def configured(db: Session, tenant_id: str | None = None) -> bool:
    return bool(api_key(db, tenant_id))


def require_key(db: Session, tenant_id: str | None = None) -> str:
    k, _, err = key_info(db, tenant_id)
    if not k:
        raise LLMNotConfigured(err or "CORTECS_API_KEY fehlt – unter Einstellungen → Geteilte Secrets setzen (oder Env CORTECS_API_KEY).")
    return k


def list_models(db: Session) -> list[dict]:
    """GET /v1/models – Rohdaten von Cortecs, auf die für die Auswahl nötigen Felder reduziert."""
    key = require_key(db)
    r = httpx.get(f"{base_url(db)}/models", headers={"Authorization": f"Bearer {key}"}, timeout=30)
    r.raise_for_status()
    data = r.json()
    items = data.get("data") if isinstance(data, dict) else data
    out = []
    for m in items or []:
        if isinstance(m, str):
            m = {"id": m}
        caps = {k: v for k, v in m.items() if isinstance(v, (bool, list, dict)) or k in ("context_size", "context_length")}
        out.append({"id": m.get("id") or m.get("name"), "raw": caps,
                    "tools": _supports_tools(m)})
    return sorted((o for o in out if o["id"]), key=lambda o: o["id"])


def _supports_tools(m: dict) -> bool | None:
    """Cortecs dokumentiert die Feldnamen nicht eindeutig → mehrere plausible Varianten prüfen; None = unbekannt."""
    flat = {}
    def walk(prefix, v):
        if isinstance(v, dict):
            for k, x in v.items():
                walk(f"{prefix}.{k}".lower(), x)
        elif isinstance(v, list):
            flat[prefix] = [str(i).lower() for i in v]
        else:
            flat[prefix] = v
    walk("", m)
    for k, v in flat.items():
        if "tool" in k or "function" in k:
            return bool(v)
        if isinstance(v, list) and any("tool" in i or "function" in i for i in v):
            return True
    return None


def summarize(db: Session, system_prompt: str, user_content: str, model: str | None = None, tenant_id: str | None = None) -> dict:
    """Ein Protokoll-Lauf wie im Core (Chat Completions), für den Prompt-Test."""
    from openai import OpenAI
    client = OpenAI(api_key=require_key(db, tenant_id), base_url=base_url(db), max_retries=2, timeout=300)
    model = model or _setting(db, "llm_model", "claude-4-6-sonnet")
    resp = client.chat.completions.create(
        model=model, temperature=float(_setting(db, "llm_temperature", "0.2")),
        max_tokens=int(_setting(db, "llm_max_tokens", "16000")), stream=False,
        messages=[{"role": "system", "content": system_prompt}, {"role": "user", "content": user_content}],
    )
    text = "\n".join(c.message.content for c in resp.choices if c.message and c.message.content).strip()
    u = getattr(resp, "usage", None)
    return {"text": text, "model": model, "input_tokens": getattr(u, "prompt_tokens", 0), "output_tokens": getattr(u, "completion_tokens", 0),
            "truncated": any(getattr(c, "finish_reason", "") in ("length", "max_tokens") for c in resp.choices)}