""" text_variants.py — 文案双轨主入口（≤100行） 轨A: generate_text_variants — 调 LLM 出 N 角度 JSON 轨B: text_import_handler — 导入外部文案进候选池 prompt 组装/解析见 _text_prompt.py；评分/去重见 text_scoring.py """ from __future__ import annotations import asyncio import logging import os from typing import Any from .constants import MAX_OPTIMIZE_ROUNDS from ._text_prompt import COPY_SYSTEM, build_prompt, parse_json_array, build_local_drafts from .text_scoring import score_copy, dedupe_copies from .llm_scorer import llm_score_copy from .banned_word_checker import check_and_fix, build_entries_from_db, CheckResult logger = logging.getLogger(__name__) async def _call_llm(client: Any, prompt: str, max_tokens: int = 8192) -> str: """统一 LLM 调用，client 由 worker 注入，隔离 key。 G1坑修复：AIClients 没有 .chat.completions，正确方法是 .chat_complete() S8: 503/429 指数退避重试（最多3次，2^attempt 秒），其他异常直接降级返 ''。 max_tokens 由调用方按批量缩放：opus 会尽量填满输出空间，8192 token 的生成 单批 >60s 必撞 apiports 网关上限返 503(task46 实测每请求恰挂 ~61s)。实测单条 max_tokens=1500~2500 仅 16~18s。故按条数动态收，墙钟压进 60s 网关窗口内。 """ import httpx # 倩倩姐2026-06-13拍板"加大重试+拉长退避"：apiports负载波动时单条opus也会被 # 拖过60s返503，短退避(1/2/4s)赶不开高负载窗口。故重试5次、退避拉长到最长30s， # 给中转站负载回落留时间。墙钟换稳定(MVP免费阶段可接受)。 max_attempts = 5 backoff = [5, 10, 20, 30] # 第1~4次重试前等待秒数，拉长跨过apiports高负载窗口 for attempt in range(max_attempts): try: return await client.chat_complete( messages=[ {"role": "system", "content": COPY_SYSTEM}, {"role": "user", "content": prompt}, ], model=client._model, max_tokens=max_tokens, temperature=0.75, ) except httpx.HTTPStatusError as exc: status = exc.response.status_code if exc.response is not None else 0 if status in (503, 429) and attempt < max_attempts - 1: wait = backoff[min(attempt, len(backoff) - 1)] logger.warning( "LLM 返回 %s，第%d/%d次重试，等待 %ds: %s", status, attempt + 1, max_attempts - 1, wait, exc, ) await asyncio.sleep(wait) continue logger.error("LLM HTTP错误(不可重试或已达上限): %s: %s", type(exc).__name__, exc) return "" except Exception as exc: # 其他异常（超时/网络断开等）不重试，直接降级 logger.error("LLM 调用失败: %s: %s", type(exc).__name__, exc) return "" return "" # apiports 网关单次响应有 ~60s 上限，claude 一次生成 >4 条长文案会超时返 503。 # 故分批：每批最多 4 条，串行调用合并。批大小可经 TEXT_BATCH_SIZE 调。 TEXT_BATCH_SIZE = int(os.environ.get("TEXT_BATCH_SIZE", "4")) async def _generate_one_batch(llm_client: Any, product: dict, batch_n: int, extra: str, strategy_narrative: str = "") -> list[dict]: """生成一批 batch_n 条，含解析重试（最多2次）。失败返回空列表。 max_tokens 按条数缩放(每条约 1800 token，封顶 8192)，压进 apiports 60s 网关窗口。""" batch_max_tokens = min(8192, max(1800, batch_n * 1800)) for attempt in range(2): raw = await _call_llm(llm_client, build_prompt( product, batch_n, extra_rules=extra, strategy_narrative=strategy_narrative, ), batch_max_tokens) parsed = parse_json_array(raw) if parsed: return parsed logger.warning("文案批(%d条)第%d次解析失败%s", batch_n, attempt + 1, "，重试" if attempt == 0 else "，放弃本批") return [] async def _generate_in_batches(llm_client: Any, product: dict, count: int, extra: str, strategy_narrative: str = "") -> list[dict]: """把 count 条按 TEXT_BATCH_SIZE 分批，串行调用合并。 串行而非并发：opus 单批就慢(~300s)且 apiports 限并发，多批 gather 会触发 大面积 503 雪崩(task45 实测)。故改串行，墙钟换稳定。""" sizes: list[int] = [] remaining = count while remaining > 0: n = min(TEXT_BATCH_SIZE, remaining) sizes.append(n) remaining -= n collected: list[dict] = [] for n in sizes: r = await _generate_one_batch(llm_client, product, n, extra, strategy_narrative) collected.extend(r) return collected async def generate_text_variants( llm_client: Any, product: dict, count: int, previous_copies: list[dict] | None = None, banned_word_rows: list[dict] | None = None, flywheel_context: str = "", strategy_narrative: str = "", ) -> list[dict]: """轨A：一次出 count 条不同角度文案，三层兜底，自动优化循环。 strategy_narrative：本套正交叙事主线(A痛点/B场景/C成分)，由调用方按套传入， 贯穿首批生成与优化轮，确保同套内文案同一叙事不串味。""" banned_entries = build_entries_from_db(banned_word_rows or []) extra = flywheel_context copies: list[dict] = await _generate_in_batches( llm_client, product, count, extra, strategy_narrative) if not copies: copies = list(build_local_drafts(product, count)) # generator → list candidates: list[dict] = [] for c in copies: ban: CheckResult = check_and_fix( f"{c.get('title','')} {c.get('content','')}", banned_entries or None, ) scored = await llm_score_copy(llm_client, c, product, [e.word for e in banned_entries]) c.update({"source": "ai", "score": scored["score"], "score_detail": scored["score_detail"], "passed": scored["passed"], "banned_word_status": ban.status, "verdict": scored.get("verdict", ""), "summary": scored.get("summary", "")}) if ban.status == "auto_fixed" and ban.fixed_text: c["content"] = ban.fixed_text candidates.append(c) failed = [c for c in candidates if not c["passed"] and c["banned_word_status"] != "hard_block"] # 优化轮默认关闭：apiports 60s 网关限制下优化轮的 _call_llm 常需白等 60s 才 503， # 严重拖慢出文案（实测 +100s+）。质量优化等北哥 prompt 方案到位再开（架构已留位）。 optimize_enabled = os.environ.get("TEXT_OPTIMIZE_ENABLED", "false").lower() == "true" rounds = MAX_OPTIMIZE_ROUNDS if optimize_enabled else 0 for _ in range(rounds): if not failed: break # 优化轮也受 60s 网关上限约束：一次最多重生成 TEXT_BATCH_SIZE 条 batch_failed = failed[:TEXT_BATCH_SIZE] hint = "\n".join( f"标题「{c['title']}」{c['score']}分，需改进：" + "；".join(d["note"] for d in c.get("score_detail", []) if d["score"] < d["max"] * 0.72) for c in batch_failed ) raw2 = await _call_llm(llm_client, build_prompt( product, len(batch_failed), extra_rules=f"以下文案未达标，请重新生成并改进：\n{hint}\n不要重复已有标题和角度。", strategy_narrative=strategy_narrative, ), min(8192, max(1800, len(batch_failed) * 1800))) if not raw2: # LLM 失败（如 503/超时）：优化是锦上添花，原始候选已够用，不再耗时重试 logger.warning("文案优化轮 LLM 失败，沿用原始候选不再重试") break for nc in parse_json_array(raw2): sc2 = await llm_score_copy(llm_client, nc, product, [e.word for e in banned_entries]) nc.update({"source": "ai", "score": sc2["score"], "score_detail": sc2["score_detail"], "passed": sc2["passed"], "banned_word_status": "pass", "verdict": sc2.get("verdict", ""), "summary": sc2.get("summary", "")}) candidates.append(nc) failed = [c for c in candidates if not c["passed"]] return dedupe_copies(candidates, previous_copies or [])[:count] def text_import_handler( raw_text: str, product: dict, banned_word_rows: list[dict] | None = None, ) -> dict: """轨B：导入外部文案（豆包等）直接进候选池，source=import""" banned_entries = build_entries_from_db(banned_word_rows or []) lines = raw_text.strip().splitlines() title = lines[0].strip() if lines else "" content = "\n".join(lines[1:]).strip() if len(lines) > 1 else raw_text.strip() candidate: dict = {"title": title, "content": content, "tags": [], "angle": "import", "buyingPoint": "", "coverTitle": title, "imageBrief": "", "source": "import"} ban = check_and_fix(f"{title} {content}", banned_entries or None) # 轨B(导入外部文案)走机械 score_copy 而非 AI 评委：导入的是用户自带成品，评分仅作 # 参考展示不卡发布；且本函数同步、改 await 会扩大到调用方。AI 评委只用于轨A生成链路。 scored = score_copy(candidate, product, [e.word for e in banned_entries]) candidate.update({"score": scored["score"], "score_detail": scored["score_detail"], "passed": scored["passed"], "banned_word_status": ban.status}) return candidate