ytrecap/server/recap.py

from __future__ import annotations

import json
import logging
import re
from datetime import datetime, timezone
from typing import Any, Literal

import httpx
from pydantic import BaseModel, Field, ValidationError, field_validator

from server.settings import settings
from server.transcripts import Transcript, format_transcript_for_prompt, transcript_hash

logger = logging.getLogger(__name__)

RECAP_SYSTEM = (
    "Du bist der Goldfish Recap Bot — trocken-witzig, ADHD-aware, nie herablassend. "
    "Du fasst YouTube-Videos für Leute zusammen, die nebenbei gespielt haben oder das Video "
    "nicht schauen wollen. Antworte NUR mit validem JSON, ohne Markdown oder Erklärungen."
)

WatchVerdict = Literal["watch", "skim", "skip"]


class RecapSection(BaseModel):
    title: str = Field(min_length=1, max_length=120)
    summary: str = Field(min_length=1, max_length=500)
    timestamp_seconds: int = Field(ge=0)
    emoji: str = Field(default="🐟", max_length=8)


class RecapPayload(BaseModel):
    tldr: str = Field(min_length=1, max_length=400)
    vibe_check: str = Field(min_length=1, max_length=200)
    watch_verdict: WatchVerdict
    goldfish_note: str = Field(min_length=1, max_length=300)
    sections: list[RecapSection] = Field(min_length=2, max_length=12)

    @field_validator("sections")
    @classmethod
    def sort_sections(cls, sections: list[RecapSection]) -> list[RecapSection]:
        return sorted(sections, key=lambda s: s.timestamp_seconds)


def utc_now_iso() -> str:
    return datetime.now(timezone.utc).replace(microsecond=0).isoformat()


def _extract_json(text: str) -> dict[str, Any]:
    cleaned = text.strip()
    if cleaned.startswith("```"):
        cleaned = re.sub(r"^```(?:json)?\s*", "", cleaned)
        cleaned = re.sub(r"\s*```$", "", cleaned)
    return json.loads(cleaned)


async def _call_openrouter(prompt: str, *, system: str = RECAP_SYSTEM) -> str | None:
    if not settings.openrouter_api_key:
        raise RuntimeError("OPENROUTER_API_KEY nicht gesetzt.")

    async with httpx.AsyncClient(timeout=120.0) as client:
        resp = await client.post(
            "https://openrouter.ai/api/v1/chat/completions",
            headers={
                "Authorization": f"Bearer {settings.openrouter_api_key}",
                "HTTP-Referer": settings.app_url,
                "X-Title": "Goldfish Recap",
            },
            json={
                "model": settings.openrouter_model,
                "messages": [
                    {"role": "system", "content": system},
                    {"role": "user", "content": prompt},
                ],
                "max_tokens": 4096,
                "response_format": {"type": "json_object"},
            },
        )
        resp.raise_for_status()
        data = resp.json()

    if data.get("error"):
        logger.error("OpenRouter error: %s", data["error"])
        return None

    choices = data.get("choices") or []
    if not choices:
        return None
    content = (choices[0].get("message") or {}).get("content")
    return str(content).strip() if content else None


def _build_prompt(
    *,
    title: str,
    channel: str,
    duration_sec: int | None,
    language: str,
    transcript_text: str,
) -> str:
    duration_hint = f"{duration_sec // 60} Minuten" if duration_sec else "unbekannt"
    return (
        f"Video: {title}\n"
        f"Kanal: {channel or 'unbekannt'}\n"
        f"Dauer: ca. {duration_hint}\n"
        f"Untertitel-Sprache: {language}\n\n"
        "Erstelle ein Recap als JSON mit exakt diesen Feldern:\n"
        "{\n"
        '  "tldr": "1-2 Sätze, das Wichtigste",\n'
        '  "vibe_check": "1 Satz Stimmung, z.B. Second-Screen-Energie 7/10",\n'
        '  "watch_verdict": "watch|skim|skip",\n'
        '  "goldfish_note": "witziger Abschluss für Goldfisch-Gedächtnis",\n'
        '  "sections": [\n'
        "    {\n"
        '      "title": "kurzer Abschnittstitel",\n'
        '      "summary": "2-3 Sätze",\n'
        '      "timestamp_seconds": 142,\n'
        '      "emoji": "🐟"\n'
        "    }\n"
        "  ]\n"
        "}\n\n"
        "Regeln:\n"
        "- 4-8 sections, timestamp_seconds MÜSSEN aus dem Transcript stammen (keine erfundenen Zeiten)\n"
        "- watch = lohnt sich ganz anzusehen, skim = nur Sprungmarken, skip = TL;DR reicht\n"
        "- Deutsch, witzig aber informativ\n\n"
        f"Transcript:\n{transcript_text}"
    )


async def generate_recap(
    *,
    metadata: dict[str, Any],
    transcript: Transcript,
    duration_sec: int | None,
) -> dict[str, Any]:
    transcript_text = format_transcript_for_prompt(transcript.segments)
    prompt = _build_prompt(
        title=metadata["title"],
        channel=metadata.get("channel") or "",
        duration_sec=duration_sec,
        language=transcript.language,
        transcript_text=transcript_text,
    )

    raw = await _call_openrouter(prompt)
    if not raw:
        raise RuntimeError("OpenRouter hat keine Antwort geliefert. Später nochmal versuchen.")

    try:
        parsed = _extract_json(raw)
        payload = RecapPayload.model_validate(parsed)
    except (json.JSONDecodeError, ValidationError) as exc:
        logger.warning("Invalid recap JSON, retrying: %s", exc)
        fix_prompt = (
            f"Fixiere dieses JSON und gib NUR valides JSON zurück:\n{raw}\n\n"
            f"Fehler: {exc}"
        )
        fixed = await _call_openrouter(fix_prompt)
        if not fixed:
            raise RuntimeError("Recap konnte nicht generiert werden.") from exc
        payload = RecapPayload.model_validate(_extract_json(fixed))

    now = utc_now_iso()
    return {
        "video_id": metadata["video_id"],
        "title": metadata["title"],
        "channel": metadata.get("channel"),
        "thumbnail_url": metadata.get("thumbnail_url"),
        "duration_sec": duration_sec,
        "language": transcript.language,
        "tldr": payload.tldr,
        "vibe_check": payload.vibe_check,
        "watch_verdict": payload.watch_verdict,
        "goldfish_note": payload.goldfish_note,
        "sections": [s.model_dump() for s in payload.sections],
        "transcript_hash": transcript_hash(transcript.segments),
        "model": settings.openrouter_model,
        "created_at": now,
        "updated_at": now,
    }