Function main

experiments/simulation/simulate_historical_episodes.py:2262–2462

()

# ==================== 主函数 ====================

def main():
    parser = argparse.ArgumentParser(description="Simulate Historical Episodes with Interest Drift")
    parser.add_argument("--start-date", type=str, required=True, help="开始日期 YYYYMMDD")
    parser.add_argument("--end-date", type=str, required=True, help="结束日期 YYYYMMDD")
    parser.add_argument(
        "--llm-model",
        type=str,
        default=os.environ.get("LLM_PARSER_OPENAI_MODEL", "gemini-3-flash-preview"),
        help="LLM model label used in benchmark summaries",
    )
    parser.add_argument("--embedding-model", type=str, default="Qwen/Qwen3-Embedding-8B", help="Embedding 模型")
    parser.add_argument("--drift-probability", type=float, default=0.5, help="漂移触发概率 (0-1)")
    parser.add_argument("--sources", nargs="*", default=None, help="每天收集的论文源 (arxiv, openreview, journal)")
    parser.add_argument("--limit-per-source", type=int, default=None, help="每天每个源的可选上限")
    parser.add_argument("--skip-paper-collection", action="store_true", help="跳过每日论文收集，只使用数据库已有论文池")
    parser.add_argument("--output-dir", type=str, default=None, help="输出目录")
    parser.add_argument("--seed", type=int, default=42, help="随机种子")
    parser.add_argument("--skip-reading-reports", action="store_true", help="Skip reading report generation for recommendation-only simulation")
    parser.add_argument("--show-count", type=int, default=DEFAULT_SIMULATION_SHOW_COUNT, help="Displayed papers per episode after real-ranking fallback fill")
    parser.add_argument("--user-count", type=int, default=None, help="Run only the first N users after stable user-id sorting")
    parser.add_argument("--user-ids", nargs="*", default=None, help="Run only these explicit user ids, e.g. user_role1 user_role9 user_role24")
    args = parser.parse_args()

    random.seed(args.seed)
    embedding_module._default_service = None
    if hasattr(reading_agent, "READING_REPORT_EVIDENCE_CACHE_ENABLED"):
        reading_agent.READING_REPORT_EVIDENCE_CACHE_ENABLED = False
    _patch_real_usage_logging()

    start_date = datetime.strptime(args.start_date, "%Y%m%d")
    end_date = datetime.strptime(args.end_date, "%Y%m%d")

    print(f"Simulating from {start_date.date()} to {end_date.date()}")
    print(f"Drift probability: {args.drift_probability}")
    print()

    # 初始化
    conn = sqlite3.connect(DB_PATH)
    users = get_all_users(conn)
    papers = get_all_papers(conn)
    conn.close()
    users = select_users(users, user_ids=args.user_ids, user_count=args.user_count)

    print(f"Loaded {len(users)} users, {len(papers)} papers")
    print(f"Selected users: {', '.join(str(user.get('user_id')) for user in users)}")

    # 加载漂移 checkfile
    checkfiles = load_checkfiles(DRIFT_CHECKFILES_DIR)
    print(f"Loaded {len(checkfiles)} drift checkfiles")

    drift_engine = DriftEngine(checkfiles)

    # 输出管理器
    output_dir = Path(args.output_dir) if args.output_dir else (PROJECT_ROOT / "data" / "simulation_output")
    if not output_dir.is_absolute():
        output_dir = PROJECT_ROOT / output_dir
    resume_state = load_resume_state(output_dir, start_date)
    if resume_state["resume"]: