""" 口播素材系统 - AI 分析脚本 用法：python analyze.py [--limit N] [--retry-failed] 功能： 1. 查询所有未分析的素材记录 2. 调用本地 oMLX（Qwen2.5-VL-3B）进行口播价值分析 3. 解析结构化 JSON 结果，写入 analysis 表 本机适配： - 使用 oMLX HTTP API（localhost:8000），非 Ollama - 模型：Qwen2.5-VL-3B-Instruct-8bit - 串行执行，防止 16GB 内存 OOM - 异常自动跳过，不中断整体流程 """ import os import sys import json import re import argparse import time sys.path.insert(0, os.path.dirname(os.path.abspath(__file__))) import httpx from utils import load_config, get_db, init_db, setup_logger # 分析提示词模板 ANALYSIS_PROMPT = """你是一个口播素材分析助手。请分析以下内容，提取口播创作可用的信息。仅返回合法JSON，不要有任何额外文字。 内容： 标题：{title} 描述：{description} 作者：{author} 平台：{platform} 要求返回JSON： {{ "tags": ["标签1","标签2","标签3"], "golden_quote": "可以直接引用的金句或观点，不超过30字", "core_idea": "一句话核心立意", "structure": "口播脚本结构：开头-中间-结尾的梗概", "emotion": "整体情绪，如 治愈/励志/震惊/沉思/有趣", "worth_downloading": "yes/no/maybe", "download_reason": "建议下载的简短理由，若no则留空" }}""" def call_omlx(prompt: str, config: dict, logger) -> str | None: """ 调用 oMLX API 进行推理 返回：模型输出的文本，失败返回 None """ import httpx try: resp = httpx.post( f"{config['llm']['base_url']}/chat/completions", headers={ "Authorization": f"Bearer {config['llm']['api_key']}", "Content-Type": "application/json", }, json={ "model": config["llm"]["model"], "messages": [{"role": "user", "content": prompt}], "response_format": {"type": "json_object"}, "temperature": config["llm"].get("temperature", 0.3), "max_tokens": 500, }, timeout=config["llm"].get("timeout", 120), ) if resp.status_code != 200: logger.error(f"oMLX 返回 {resp.status_code}: {resp.text[:200]}") return None return resp.json()["choices"][0]["message"]["content"] except httpx.TimeoutException: logger.error("oMLX 请求超时") return None except Exception as e: logger.error(f"oMLX 调用异常: {e}") return None def parse_analysis_json(raw_text: str, logger) -> dict | None: """ 解析模型输出的 JSON，带容错处理 """ if not raw_text: return None # 尝试直接解析 try: result = json.loads(raw_text) if isinstance(result, dict): return result except json.JSONDecodeError: pass # 尝试提取 JSON 块（模型可能输出 ```json ... ```） match = re.search(r'```(?:json)?\s*(\{.*?\})\s*```', raw_text, re.DOTALL) if match: try: result = json.loads(match.group(1)) if isinstance(result, dict): return result except json.JSONDecodeError: pass # 尝试提取第一个 { 到最后一个 } first = raw_text.find("{") last = raw_text.rfind("}") if first != -1 and last != -1 and last > first: try: result = json.loads(raw_text[first:last+1]) if isinstance(result, dict): return result except json.JSONDecodeError: pass logger.warning(f"JSON 解析失败，原始输出: {raw_text[:200]}") return None def analyze_one(material: dict, config: dict, logger) -> dict | None: """ 分析单条素材 """ title = material["title"] or "" description = (material.get("description") or "")[:500] # 截断避免过长 if not title and not description: logger.warning(f"素材 {material['id']} 无标题和描述，跳过") return None prompt = ANALYSIS_PROMPT.format( title=title, description=description, author=material.get("author", "未知"), platform=material.get("platform", "未知"), ) raw_output = call_omlx(prompt, config, logger) if not raw_output: return None result = parse_analysis_json(raw_output, logger) if not result: return None # 规范化字段 return { "material_id": material["id"], "tags": ",".join(result.get("tags", [])), "golden_quote": str(result.get("golden_quote", ""))[:200], "core_idea": str(result.get("core_idea", ""))[:300], "structure": str(result.get("structure", ""))[:500], "emotion": str(result.get("emotion", ""))[:20], "worth_downloading": str(result.get("worth_downloading", "no"))[:10], "download_reason": str(result.get("download_reason", ""))[:200], "raw_model_output": raw_output, } def main(): parser = argparse.ArgumentParser(description="口播素材 AI 分析") parser.add_argument("--limit", "-n", type=int, default=0, help="分析数量上限，0=全部") parser.add_argument("--retry-failed", action="store_true", help="重试分析失败的记录") parser.add_argument("--dry-run", action="store_true", help="仅显示待分析数量，不执行") args = parser.parse_args() config = load_config() logger = setup_logger("analyze") # 检查 oMLX from utils import check_omlx if not check_omlx(config): logger.error("oMLX 不可用，请先启动 oMLX 服务") return db_path = init_db() db = get_db(db_path) # 查询待分析记录 if args.retry_failed: query = "SELECT * FROM materials WHERE analyzed = 2" else: query = "SELECT * FROM materials WHERE analyzed = 0" if args.limit > 0: query += f" LIMIT {args.limit}" materials = db.execute(query).fetchall() materials = [dict(m) for m in materials] if not materials: print("[OK] 没有待分析的素材") db.close() return print(f"[INFO] 待分析: {len(materials)} 条") if args.dry_run: for m in materials[:5]: print(f" - [{m['platform']}] {m['title'][:40]}") if len(materials) > 5: print(f" ... 还有 {len(materials) - 5} 条") db.close() return success = 0 failed = 0 for i, material in enumerate(materials, 1): print(f"\r[进度] {i}/{len(materials)} 分析中...", end="", flush=True) try: result = analyze_one(material, config, logger) if result: # 写入 analysis 表 db.execute(""" INSERT OR REPLACE INTO analysis (material_id, tags, golden_quote, core_idea, structure, emotion, worth_downloading, download_reason, raw_model_output) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?) """, ( result["material_id"], result["tags"], result["golden_quote"], result["core_idea"], result["structure"], result["emotion"], result["worth_downloading"], result["download_reason"], result["raw_model_output"], )) # 更新素材分析状态 db.execute("UPDATE materials SET analyzed = 1 WHERE id = ?", (material["id"],)) db.commit() success += 1 else: db.execute("UPDATE materials SET analyzed = 2 WHERE id = ?", (material["id"],)) db.commit() failed += 1 # 间隔 1 秒，避免过载 time.sleep(1) except Exception as e: logger.error(f"分析异常 [{material['id']}]: {e}") db.execute("UPDATE materials SET analyzed = 2 WHERE id = ?", (material["id"],)) db.commit() failed += 1 db.close() print(f"\r[DONE] 分析完成: 成功 {success}, 失败 {failed}, 共 {len(materials)} 条") if __name__ == "__main__": main()