xiaoxia-saas/packages/shared/ai_service.py

"""AI 服务层 — 智能推荐 & 封面生成.

提供 AI 推荐片段编排方案和封面生成的核心业务逻辑。
API 层和 Worker 层都从此模块导入，避免 API 直接依赖 Worker 代码。
"""

from __future__ import annotations

import copy
import json
import logging
import random
import time
from typing import Any, Dict, List, Optional

from packages.domain.config_schemas import DEFAULT_EDIT_PLAN_CONFIG
from packages.shared.ai_client import get_doubao_client

logger = logging.getLogger(__name__)


# ── AI 推荐片段方案 ──────────────────────────────────────────────────────────


def _fallback_recommend_clips(
    plan_id: str,
    template_id: str,
    asset_ids: List[str],
    editing_mode: str,
    target_duration: float,
) -> Dict[str, Any]:
    """本地降级推荐方案（原 stub 逻辑）.

    当豆包 API 不可用或调用失败时使用，基于模板规则生成模拟推荐数据。
    """
    # 模拟 AI 分析耗时
    time.sleep(0.5)

    # 根据素材数量生成推荐片段
    clips: List[Dict[str, Any]] = []
    order = 0

    # 开场片段
    clips.append(
        {
            "clip_type": "intro",
            "order": order,
            "text_content": "精彩看点",
            "duration": 3.0,
            "transition_effect": "fade",
            "asset_id": asset_ids[0] if asset_ids else "",
            "start_time": 0.0,
            "config": {},
        }
    )
    order += 1

    # 为每个素材生成展示片段
    per_clip_duration = max(2.0, (target_duration - 6.0) / max(len(asset_ids), 1))
    for i, asset_id in enumerate(asset_ids):
        clips.append(
            {
                "clip_type": "showcase",
                "order": order,
                "text_content": f"展示片段 {i + 1}",
                "duration": round(per_clip_duration, 1),
                "transition_effect": "cut",
                "asset_id": asset_id,
                "start_time": 0.0,
                "config": {},
            }
        )
        order += 1

    # 结尾 CTA
    clips.append(
        {
            "clip_type": "outro",
            "order": order,
            "text_content": "感谢观看",
            "duration": 3.0,
            "transition_effect": "fade",
            "asset_id": "",
            "start_time": 0.0,
            "config": {},
        }
    )
    order += 1

    # 生成推荐 config
    config = copy.deepcopy(DEFAULT_EDIT_PLAN_CONFIG)
    config["title"]["text"] = f"精选视频 — {len(asset_ids)} 个片段"
    config["title"]["ai_auto"] = True

    return {
        "clips": clips,
        "config": config,
        "total_duration": round(sum(c["duration"] for c in clips), 1),
        "confidence": round(random.uniform(0.75, 0.95), 2),
    }


def _parse_recommend_response(
    content: str,
    asset_ids: List[str],
    target_duration: float,
) -> Optional[Dict[str, Any]]:
    """解析豆包返回的推荐方案.

    期望返回结构：
    {
        "clips": [
            {"clip_type": "intro/showcase/outro", "order": 0,
             "text_content": "...", "duration": 3.0,
             "transition_effect": "fade/cut", "asset_id": "...",
             "start_time": 0.0, "config": {}}
        ],
        "title": "视频标题",
        "confidence": 0.85
    }
    """
    if not content:
        return None

    try:
        cleaned = content.strip()
        if cleaned.startswith("```"):
            cleaned = cleaned.strip("`")
            if cleaned.lower().startswith("json"):
                cleaned = cleaned[4:]
            cleaned = cleaned.strip()

        data = json.loads(cleaned)
        if not isinstance(data, dict):
            return None

        clips_data = data.get("clips", [])
        if not isinstance(clips_data, list) or len(clips_data) == 0:
            return None

        clips: List[Dict[str, Any]] = []
        for _, clip in enumerate(clips_data):
            if not isinstance(clip, dict):
                continue
            asset_id = str(clip.get("asset_id", ""))
            # 校验 asset_id 是否在输入列表中
            if asset_id and asset_id not in asset_ids:
                asset_id = ""
            clips.append(
                {
                    "clip_type": clip.get("clip_type", "showcase"),
                    "order": clip.get("order", len(clips)),
                    "text_content": str(clip.get("text_content", "")),
                    "duration": max(1.0, min(30.0, float(clip.get("duration", 3.0)))),
                    "transition_effect": clip.get("transition_effect", "cut"),
                    "asset_id": asset_id,
                    "start_time": max(0.0, float(clip.get("start_time", 0.0))),
                    "config": clip.get("config", {}) or {},
                }
            )

        if not clips:
            return None

        # 按 order 排序
        clips.sort(key=lambda c: c["order"])
        # 重新编号 order 保证连续
        for i, clip in enumerate(clips):
            clip["order"] = i

        config = copy.deepcopy(DEFAULT_EDIT_PLAN_CONFIG)
        title = data.get("title", "")
        if title:
            config["title"]["text"] = str(title)
            config["title"]["ai_auto"] = True

        confidence = float(data.get("confidence", 0.7))
        confidence = max(0.0, min(1.0, confidence))

        total_duration = round(sum(c["duration"] for c in clips), 1)

        return {
            "clips": clips,
            "config": config,
            "total_duration": total_duration,
            "confidence": round(confidence, 2),
        }

    except (json.JSONDecodeError, ValueError, TypeError, KeyError):
        return None


def _call_ai_recommend_service(
    plan_id: str,
    template_id: str,
    asset_ids: List[str],
    editing_mode: str,
    target_duration: float,
    asset_analyses: Optional[Dict[str, str]] = None,
) -> Dict[str, Any]:
    """调用 AI 推荐服务生成片段编排方案.

    优先使用豆包大模型生成，失败或未配置时降级为本地规则生成。
    当提供 asset_analyses 时，会将每个素材的视频理解结果注入 prompt，
    让 LLM 能基于视频实际内容做智能编排。

    Args:
        plan_id: 剪辑计划 ID
        template_id: 模板 ID
        asset_ids: 素材 ID 列表
        editing_mode: 剪辑模式
        target_duration: 目标时长（秒）
        asset_analyses: 可选，{asset_id: 视频理解文本} 映射
    """
    client = get_doubao_client()
    if not client.is_available:
        logger.info("豆包API未配置，使用本地降级生成AI推荐方案")
        return _fallback_recommend_clips(plan_id, template_id, asset_ids, editing_mode, target_duration)

    # 构建素材描述（含视频理解结果）
    asset_analyses = asset_analyses or {}
    asset_lines = []
    for aid in asset_ids[:30]:
        analysis = asset_analyses.get(aid, "")
        if analysis:
            # 截断过长的分析结果，避免 token 爆炸
            analysis_truncated = analysis[:300] + ("..." if len(analysis) > 300 else "")
            asset_lines.append(f"  - 素材ID: {aid}\n    内容描述: {analysis_truncated}")
        else:
            asset_lines.append(f"  - 素材ID: {aid}")

    assets_desc = "\n".join(asset_lines)
    has_analysis = any(aid in asset_analyses for aid in asset_ids[:30])

    # 构建 prompt
    system_prompt = (
        "你是一个专业的视频剪辑导演助手。" "根据提供的素材列表和目标时长，设计一个完整的视频片段编排方案。\n"
    )
    if has_analysis:
        system_prompt += (
            "每个素材附带了 AI 视频理解的内容描述，请根据素材的实际内容来决策编排：\n"
            "- 将内容相关的素材放在一起，保持叙事连贯\n"
            "- 根据素材内容合理安排片段顺序（如开场用吸引人的画面、高潮部分紧凑切换等）\n"
            "- 为每个片段选择最匹配的素材，并在 text_content 中体现素材主题\n"
        )
    system_prompt += (
        "要求：\n"
        "1. 片段类型分为三类：intro（开场）、showcase（展示）、outro（结尾）\n"
        "2. 每个片段包含：clip_type、order、text_content（字幕/标题文字）、"
        "duration（时长秒）、transition_effect（转场效果：fade/cut/dissolve）、"
        "asset_id（使用的素材ID）、start_time（素材起始时间秒）\n"
        "3. 总时长接近 target_duration，每个素材至少用一次\n"
        "4. 转场效果合理分配，不要全用cut\n"
        "5. 返回纯JSON，不要其他文字\n"
        '返回格式：{"clips": [...], "title": "视频标题", "confidence": 0.85}'
    )

    user_prompt = (
        f"剪辑计划ID: {plan_id}\n"
        f"模板ID: {template_id}\n"
        f"剪辑模式: {editing_mode}\n"
        f"目标时长: {target_duration}秒\n"
        f"素材列表（共{len(asset_ids)}个）：\n{assets_desc}\n\n"
        f"请设计完整的片段编排方案："
    )

    messages = [
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": user_prompt},
    ]

    result = client.chat_completion(
        messages=messages,
        temperature=0.7,
        max_tokens=2048,
    )

    if result:
        parsed = _parse_recommend_response(result, asset_ids, target_duration)
        if parsed and len(parsed["clips"]) >= 2:
            logger.info(
                "豆包AI推荐生成成功: plan_id=%s clips=%d duration=%.1f confidence=%.2f has_analysis=%s",
                plan_id,
                len(parsed["clips"]),
                parsed["total_duration"],
                parsed["confidence"],
                has_analysis,
            )
            return parsed
        logger.warning("豆包AI推荐返回解析失败，降级到本地方案: %s", result[:100])

    # 降级
    return _fallback_recommend_clips(plan_id, template_id, asset_ids, editing_mode, target_duration)


# ── AI 封面生成 ──────────────────────────────────────────────────────────────


def _call_ai_cover_service(
    plan_id: str,
    asset_ids: List[str],
    cover_type: str,
    frame_time: float | None = None,
    primary_video_url: str | None = None,
) -> Dict[str, Any]:
    """调用 AI 封面生成服务.

    当 cover_type 为 ai_frame 或 ai_regenerate 时，调用 MediaKit 视频截帧。
    失败或未配置时降级为 stub 行为。

    Args:
        plan_id: 剪辑计划 ID
        asset_ids: 素材 ID 列表
        cover_type: 封面类型
        frame_time: 手动选帧时间点
        primary_video_url: 主视频的可访问 URL（用于 MediaKit 抽帧）
    """
    if cover_type == "upload":
        return {
            "type": "upload",
            "image_url": "",
            "message": "请上传封面图片",
        }

    if cover_type == "manual" and frame_time is not None:
        return {
            "type": "manual",
            "image_url": f"/api/v1/assets/placeholder/cover?time={frame_time}",
            "frame_time": frame_time,
        }

    # ai_frame / ai_regenerate - 尝试调用 MediaKit
    if primary_video_url:
        from packages.shared.mediakit_client import get_mediakit_client

        client = get_mediakit_client()
        if client.is_available:
            try:
                logger.info("调用 MediaKit 抽帧: plan_id=%s video=%s", plan_id, primary_video_url[:80])
                frames = client.extract_frames(
                    video_url=primary_video_url,
                    strategy="TimeInterval",
                    max_frames=5,
                )

                if frames and len(frames) > 0:
                    # 选择第一帧（SceneChange 策略的第一帧通常是最佳画面）
                    best_frame = frames[0]
                    image_url = best_frame.get("image_url", "")
                    timestamp = best_frame.get("timestamp", 0.0)

                    if image_url:
                        logger.info(
                            "MediaKit 抽帧成功: plan_id=%s frame_time=%.2f url=%s",
                            plan_id,
                            timestamp,
                            image_url[:80],
                        )
                        return {
                            "type": "ai_frame",
                            "image_url": image_url,
                            "frame_time": round(timestamp, 1),
                            "confidence": 0.85,
                        }
                    else:
                        logger.warning("MediaKit 返回的帧无 image_url")

            except Exception as e:
                logger.exception("MediaKit 抽帧失败，降级到 stub: %s", str(e))

    # 降级：stub 行为
    logger.info("使用 stub 封面: plan_id=%s", plan_id)
    time.sleep(0.3)
    return {
        "type": "ai_frame",
        "image_url": f"/api/v1/assets/placeholder/cover?plan={plan_id}",
        "frame_time": round(random.uniform(1.0, 10.0), 1),
        "confidence": round(random.uniform(0.80, 0.98), 2),
    }


# ── 公共入口 ────────────────────────────────────────────────────────────────


def run_ai_recommend(
    plan_id: str,
    template_id: str,
    asset_ids: List[str],
    editing_mode: str = "one_take",
    target_duration: float = 30.0,
    asset_analyses: Optional[Dict[str, str]] = None,
) -> Dict[str, Any]:
    """执行 AI 推荐片段方案

    Args:
        plan_id: 剪辑计划 ID
        template_id: 模板 ID
        asset_ids: 素材 ID 列表
        editing_mode: 剪辑模式 (one_take / pip / voice_over / voice_pip)
        target_duration: 目标时长（秒）
        asset_analyses: 可选，{asset_id: 视频理解文本} 映射

    Returns:
        推荐方案 dict，包含 clips / config / total_duration / confidence
    """
    logger.info(
        "AI 推荐片段方案: plan_id=%s template_id=%s assets=%d mode=%s duration=%.1f has_analysis=%s",
        plan_id,
        template_id,
        len(asset_ids),
        editing_mode,
        target_duration,
        bool(asset_analyses),
    )
    result = _call_ai_recommend_service(
        plan_id=plan_id,
        template_id=template_id,
        asset_ids=asset_ids,
        editing_mode=editing_mode,
        target_duration=target_duration,
        asset_analyses=asset_analyses,
    )
    logger.info(
        "AI 推荐完成: plan_id=%s clips=%d duration=%.1f confidence=%.2f",
        plan_id,
        len(result["clips"]),
        result["total_duration"],
        result["confidence"],
    )
    return result


def run_generate_cover(
    plan_id: str,
    asset_ids: List[str],
    cover_type: str = "ai_frame",
    frame_time: float | None = None,
    primary_video_url: str | None = None,
) -> Dict[str, Any]:
    """执行 AI 封面生成

    Args:
        plan_id: 剪辑计划 ID
        asset_ids: 素材 ID 列表（用于确定视频来源）
        cover_type: 封面类型 (ai_frame / manual / upload / ai_regenerate)
        frame_time: 手动选帧时间点（仅 manual 模式使用）
        primary_video_url: 主视频的可访问 URL（用于 MediaKit 抽帧）

    Returns:
        封面数据 dict，包含 type / image_url / frame_time
    """
    logger.info(
        "AI 封面生成: plan_id=%s type=%s assets=%d has_video_url=%s",
        plan_id,
        cover_type,
        len(asset_ids),
        bool(primary_video_url),
    )
    result = _call_ai_cover_service(
        plan_id=plan_id,
        asset_ids=asset_ids,
        cover_type=cover_type,
        frame_time=frame_time,
        primary_video_url=primary_video_url,
    )
    logger.info(
        "AI 封面生成完成: plan_id=%s type=%s url=%s",
        plan_id,
        result.get("type"),
        result.get("image_url", "")[:60],
    )
    return result