From 90004cced4200edde445e1dfe4ca8688ae14e509 Mon Sep 17 00:00:00 2001 From: xiaoxia Date: Mon, 5 Oct 2026 20:03:30 +0800 Subject: [PATCH] refactor(vision): dashscope-only (#2207) --- apps/worker/worker_app/tasks/viral_video.py | 30 +- .../worker_app/tasks/vision/fast_path.py | 31 +- .../worker_app/tasks/vision/vlm_fallback.py | 302 +++++++----------- .../worker_app/tasks/vision/vlm_fast_json.py | 179 ++++------- 4 files changed, 212 insertions(+), 330 deletions(-) diff --git a/apps/worker/worker_app/tasks/viral_video.py b/apps/worker/worker_app/tasks/viral_video.py index 68155919e..436a5d481 100644 --- a/apps/worker/worker_app/tasks/viral_video.py +++ b/apps/worker/worker_app/tasks/viral_video.py @@ -3,7 +3,7 @@ V2 图片分析（10-05）：火山OCR专用API + doubao-lite强约束JSON并行，单图<3s，8图<15s；pro VLM单次兜底。输出字段兼容旧格式，下游信任链/t2i零改动。 流水线步骤： - 1. _step_image_analysis 图片分析（V2: OCR+lite VLM并行 + pro兜底） + 1. _step_image_analysis 图片分析（V2: OCR+qwen3.8-flash并行 + qwen3.7-plus兜底） 1.5 _step_video_analysis 参考视频风格分析（可选） 2. _step_intent_parsing 用户文案意图解析 3. _step_script_generation 编导分镜脚本生成（融合原 copy_fusion+storyboard+review，输出 copy_result 结构 + voiceover_script） @@ -358,10 +358,10 @@ def _step_image_analysis(job: ViralVideoJob) -> dict: """步骤 1: 图片分析（V2 主路径）。 架构： - - 主力：火山 MediaKit OCR（专用API）+ doubao-seed-2.1-lite 强约束 JSON（弥补火山云端缺失的 - 人体属性/商品检测/图像标签专用HTTP API），每图2路并行，目标<3s； + - 主力：火山 MediaKit OCR（专用API，未配置时自动跳过）+ qwen3.8-flash 强约束 JSON，每图2路并行，目标<3s； - 外层全并发（workers=8），目标8图<15s； - - 兜底：fast 结果不可用时单次调用 doubao-seed-2.1-pro VLM（简单、无竞速）。 + - 兜底：fast 结果不可用时单次调用 qwen3.7-plus（简单、无竞速）。 + - 唯一后端：阿里云百炼 DashScope，API Key 从环境变量 DASHSCOPE_API_KEY 读取。 输出 dict 字段（name/brand/category/appearance/key_features/scene/mood/portrait_prompt/summary/_source） 与旧版格式完全一致，下游信任链/t2i/intent_parsing/script_generation 零改动。 """ @@ -383,26 +383,8 @@ def _step_image_analysis(job: ViralVideoJob) -> dict: logger.error("[爆款视频] vision 模块导入失败: %s", e) return {"products": [_vision_fallback(0, f"vision_import_error:{e}")]} - # 整个阶段关闭底层 httpx 重试，避免线程里出现不可控等待 - try: - from packages.shared.ai_client import get_doubao_client as _gdc - - _cli = _gdc() - _orig_retries = _cli.max_retries - _cli.max_retries = 0 - except Exception: - _cli = None - _orig_retries = 0 - - try: - results = _aiv2(normalized_urls) - finally: - if _cli is not None: - try: - _cli.max_retries = _orig_retries - except Exception: - pass - + # V2 内部 httpx 直连 dashscope，单次调用无重试，无需调整全局 client + results = _aiv2(normalized_urls) return {"products": list(results)} diff --git a/apps/worker/worker_app/tasks/vision/fast_path.py b/apps/worker/worker_app/tasks/vision/fast_path.py index 323ee44ea..e05ca4e4c 100644 --- a/apps/worker/worker_app/tasks/vision/fast_path.py +++ b/apps/worker/worker_app/tasks/vision/fast_path.py @@ -1,9 +1,11 @@ # -*- coding: utf-8 -*- -"""V2 图片分析主路径：每图并行 OCR（火山专用API）+ lite JSON VLM，失败时单次 pro VLM 兜底。 +"""V2 图片分析主路径：每图并行 OCR（火山MediaKit，未配置时自动跳过）+ qwen3.8-flash JSON VLM， +失败时单次 qwen3.7-plus 兜底。 -设计原则（灵应10-05要求）： -- 主力路径简洁：单图2路并行，外层N图全并发 -- 兜底简单：单次 pro VLM 调用，无竞速/重试/复杂超时 +架构（灵应10-05确认）： +- 唯一后端：阿里云百炼 DashScope，qwen3.8-flash 做快速路径、qwen3.7-plus 做兜底 +- 主力：单图2路并行（OCR + fast VLM），外层N图全并发（workers=8） +- 兜底：单次 pro VLM 调用，无竞速/重试/复杂超时 - 输出 dict 格式与旧版完全一致，下游零改动 """ @@ -19,12 +21,12 @@ from . import assembler, ocr_volc, vlm_fallback, vlm_fast_json logger = logging.getLogger(__name__) -# 可通过环境变量调参（有默认值，无需配置即可跑） +# 超时（可通过环境变量覆盖） _IMG_WORKERS = int(os.environ.get("VISION_V2_IMG_WORKERS", "8")) _FAST_TIMEOUT = float(os.environ.get("VISION_V2_FAST_TIMEOUT", "8")) _FAST_JSON_TIMEOUT = float(os.environ.get("VISION_V2_FAST_JSON_TIMEOUT", "8")) _OCR_TIMEOUT = float(os.environ.get("VISION_V2_OCR_TIMEOUT", "6")) -_PRO_TIMEOUT = float(os.environ.get("VISION_V2_PRO_TIMEOUT", "45")) +_PRO_TIMEOUT = float(os.environ.get("VISION_V2_PRO_TIMEOUT", "20")) _FALLBACK_RESULT = { "name": "未识别", @@ -42,7 +44,6 @@ _FALLBACK_RESULT = { def _is_usable(r: dict[str, Any]) -> bool: - """结果可用判定：portrait_prompt 是核心，有效就算 usable。""" pp = (r.get("portrait_prompt") or "").strip() if pp and pp not in ("无人像", "无法判断", "未识别"): return True @@ -53,10 +54,8 @@ def _is_usable(r: dict[str, Any]) -> bool: def analyze_image_v2(idx: int, img_url: str) -> dict[str, Any]: - """单张图片 V2 分析。""" t0 = time.time() - # 第1层：OCR + lite JSON VLM 并行 fj_result: dict[str, Any] | None = None ocr_result: list[str] = [] with ThreadPoolExecutor(max_workers=2) as pool: @@ -74,7 +73,6 @@ def analyze_image_v2(idx: int, img_url: str) -> dict[str, Any]: elif fut is f_ocr and isinstance(res, list): ocr_result = res except TimeoutError: - # fast 整体超时，取消还没跑完的子任务，继续走 pro 兜底 for f in (f_fj, f_ocr): if not f.done(): f.cancel() @@ -82,20 +80,16 @@ def analyze_image_v2(idx: int, img_url: str) -> dict[str, Any]: fast_elapsed = time.time() - t0 - # 组装 fast 结果 if fj_result: assembled = assembler.assemble_result(idx, fj_result, ocr_result) if _is_usable(assembled): assembled["_fast_elapsed"] = round(fast_elapsed, 2) logger.info( "[vision.v2] 图片 #%d fast命中 elapsed=%.2fs pp=%s", - idx, - fast_elapsed, - (assembled.get("portrait_prompt") or "")[:40], + idx, fast_elapsed, (assembled.get("portrait_prompt") or "")[:40], ) return assembled - # 第2层：pro VLM 单次兜底 pro_t0 = time.time() pro_result = vlm_fallback.call_pro_vlm(img_url, idx, timeout=_PRO_TIMEOUT) if pro_result and _is_usable(pro_result): @@ -107,7 +101,6 @@ def analyze_image_v2(idx: int, img_url: str) -> dict[str, Any]: logger.info("[vision.v2] 图片 #%d pro兜底命中 total=%.2fs", idx, time.time() - t0) return pro_result - # 最终：返回最小可用结果 logger.warning("[vision.v2] 图片 #%d 全路径失败 elapsed=%.2fs", idx, time.time() - t0) out = dict(_FALLBACK_RESULT) out["_source"] = "v2_all_failed" @@ -117,13 +110,15 @@ def analyze_image_v2(idx: int, img_url: str) -> dict[str, Any]: def analyze_images_v2(img_urls: list[str]) -> list[dict[str, Any]]: - """批量图片 V2 分析，外层全并发。""" if not img_urls: return [] workers = min(_IMG_WORKERS, len(img_urls), 16) results: list[dict[str, Any] | None] = [None] * len(img_urls) - logger.info("[vision.v2] 开始图片分析 n=%d workers=%d fast_timeout=%.0fs", len(img_urls), workers, _FAST_TIMEOUT) + logger.info( + "[vision.v2] 开始图片分析 n=%d workers=%d fast_timeout=%.0fs pro_timeout=%.0fs", + len(img_urls), workers, _FAST_TIMEOUT, _PRO_TIMEOUT, + ) t0 = time.time() with ThreadPoolExecutor(max_workers=workers) as pool: future_to_idx = {pool.submit(analyze_image_v2, idx, url): idx for idx, url in enumerate(img_urls)} diff --git a/apps/worker/worker_app/tasks/vision/vlm_fallback.py b/apps/worker/worker_app/tasks/vision/vlm_fallback.py index 558500c17..552660d8b 100644 --- a/apps/worker/worker_app/tasks/vision/vlm_fallback.py +++ b/apps/worker/worker_app/tasks/vision/vlm_fallback.py @@ -1,22 +1,46 @@ # -*- coding: utf-8 -*- -"""VLM 兜底：专用API路径失败时的最后一道防线，单次调用 doubao-seed-2.1-pro。 +"""V2 pro 兜底：qwen3.7-plus（阿里云百炼/DashScope）单次调用。 -设计原则：简单、直接、无竞速、无复杂超时逻辑。只在 fast_json 结果不可用时调用。 +fast_json 结果不可用时单次调用，无竞速、无重试、无复杂超时逻辑。 +直接 httpx 发精简 JSON-only prompt（比旧版 prompt_loader XML 模板短很多，降低延迟）。 """ from __future__ import annotations import json import logging -import re +import os import time from typing import Any logger = logging.getLogger(__name__) -DEFAULT_PRO_MODEL = "doubao-seed-2-1-pro-260915" -DEFAULT_TIMEOUT = 45 -DEFAULT_MAX_TOKENS = 800 +_BASE_URL = "https://dashscope.aliyuncs.com/compatible-mode/v1" +_PRO_MODEL = "qwen3.7-plus" +_DEFAULT_TIMEOUT = 20 +_DEFAULT_MAX_TOKENS = 800 + +_PRO_SYSTEM = ( + "你是图片分析助手。仔细观察图片，严格按JSON schema返回一个对象，不要任何解释、" + "不要markdown、不要代码块、不要前后缀文字。字段值不确定时填null或空数组。\n" + "{\n" + ' "has_person": true/false,\n' + ' "gender": "男"/"女"/null,\n' + ' "age_range": "儿童"/"青少年"/"青年"/"中年"/"老年"/null,\n' + ' "outfit": "人物穿搭描述，60字以内（例：白色T恤+牛仔裤）",\n' + ' "hair": "发型",\n' + ' "pose": "姿态",\n' + ' "expression": "表情",\n' + ' "scene": "场景",\n' + ' "mood": "氛围",\n' + ' "has_product": true/false,\n' + ' "category": "服饰/鞋包/美妆/数码/食品/家居/配饰/母婴/非产品图",\n' + ' "product_name": "产品名称，非产品图填null",\n' + ' "brand": "品牌或文字标识，无则null",\n' + ' "key_features": ["特征数组"]\n' + "}" +) +_PRO_USER = "分析这张图片，返回符合schema的JSON。" def _strip_code_fence(s: str) -> str: @@ -31,196 +55,114 @@ def _strip_code_fence(s: str) -> str: return s -def _xml_text(tag: str, xml: str) -> str: - m = re.search(rf"<{tag}[^>]*>(.*?)", xml, re.S) - return (m.group(1) if m else "").strip() - - -def _xml_attr(tag: str, attr: str, xml: str) -> str: - m = re.search(rf"<{tag}[^>]*\b{attr}\s*=\s*[\"']([^\"']*)[\"']", xml) - return (m.group(1) if m else "").strip() - - -def _xml_to_product(raw: str, idx: int) -> dict[str, Any]: - """解析 VLM 输出的 XML 格式（简化版）。""" - scene = _xml_text("scene", raw) or "通用" - mood = _xml_text("mood", raw) or "" - - portrait_prompt = "无人像" - p_has = _xml_attr("people", "has_person", raw) - if p_has and p_has.lower() != "false": - gender = _xml_attr("people", "gender", raw) or "" - age = _xml_attr("people", "age_range", raw) or "" - outfit = _xml_attr("people", "outfit", raw) or "" - hair = _xml_attr("people", "hair", raw) or "自然发型" - pose = _xml_attr("people", "pose", raw) or "" - expr = _xml_attr("people", "expression", raw) or "自然" - parts: list[str] = [] - if gender: - parts.append(gender + ("性" if not gender.endswith("性") else "")) - if age: - parts.append(age) - parts.append("人物") +def _assemble_pp(obj: dict[str, Any]) -> str: + if not obj.get("has_person", False): + return "无人像" + parts: list[str] = [] + gender = obj.get("gender") + age = obj.get("age_range") + if gender: + parts.append(gender + ("性" if not gender.endswith("性") else "")) + if age: + parts.append(age) + parts.append("人物") + hair = obj.get("hair") + if hair: parts.append(hair) - if outfit: - parts.append(f"身着{outfit}") - if pose: - parts.append(f"姿态{pose}") + outfit = obj.get("outfit") + if outfit: + parts.append(f"身着{outfit}") + pose = obj.get("pose") + if pose: + parts.append(f"姿态{pose}") + expr = obj.get("expression") + if expr: parts.append(f"表情{expr}") - portrait_prompt = "，".join(parts) - - m = re.search(r"]*>(.*?)", raw, re.S) - if m: - pbody = m.group(1) - name = _xml_attr("product", "name", raw) or _xml_text("name", pbody) or "未识别" - brand = _xml_attr("product", "brand", raw) or _xml_text("brand", pbody) or "无法判断" - category = _xml_attr("product", "category", raw) or _xml_text("category", pbody) or "无法判断" - appearance = _xml_attr("product", "appearance", raw) or _xml_text("appearance", pbody) or "无法判断" - packaging = _xml_attr("product", "packaging", raw) or _xml_text("packaging", pbody) or "无法判断" - feat = _xml_attr("product", "features", raw) or _xml_text("features", pbody) or "" - feat_list = [x.strip() for x in re.split(r"[,，;；]", feat) if x.strip()] if feat else ["无法判断"] - top_text = _xml_attr("product", "text_on_package", raw) or _xml_text("text_on_package", pbody) or "" - text_list = [x.strip() for x in re.split(r"[,，;；]", top_text) if x.strip()] if top_text else [] - summary = _xml_attr("product", "summary", raw) or _xml_text("summary", pbody) or f"{brand} {name}" - pp_attr = _xml_attr("product", "portrait_prompt", raw) - if pp_attr and pp_attr != "无人像": - portrait_prompt = pp_attr - return { - "name": name, - "brand": brand, - "category": category, - "appearance": appearance, - "packaging": packaging, - "text_on_package": text_list, - "key_features": feat_list, - "scene": scene, - "mood": mood, - "portrait_prompt": portrait_prompt, - "summary": summary, - "_source": "vlm_pro_xml", - } - - if portrait_prompt != "无人像": - return { - "name": "未识别", - "brand": "无法判断", - "category": "无法判断", - "appearance": "无法判断", - "packaging": "无法判断", - "text_on_package": [], - "key_features": ["无法判断"], - "scene": scene, - "mood": mood, - "portrait_prompt": portrait_prompt, - "summary": "未识别", - "_source": "vlm_pro_no_product", - } - return { - "name": "未识别", - "brand": "无法判断", - "category": "无法判断", - "appearance": "无法判断", - "packaging": "无法判断", - "text_on_package": [], - "key_features": ["无法判断"], - "scene": scene, - "mood": mood, - "portrait_prompt": "无人像", - "summary": "未识别", - "_source": "vlm_pro_no_tag", - } + return "，".join(parts) if parts else "无人像" def call_pro_vlm( img_url: str, idx: int, *, - model: str | None = None, - timeout: int = DEFAULT_TIMEOUT, + timeout: int = _DEFAULT_TIMEOUT, ) -> dict[str, Any] | None: - """单次调用 pro VLM，解析后返回 product dict；失败返回 None。""" + """单次调用 qwen3.7-plus，解析后返回 product dict；失败返回 None。""" t0 = time.time() - try: - from packages.application.viral_video.prompt_loader import ( - get_template, - render_system_prompt, - render_user_prompt, - ) - from packages.shared.ai_client import get_doubao_client - except ImportError as e: - logger.warning("[vision.vlm] 导入失败: %s", e) + import httpx + + api_key = os.environ.get("DASHSCOPE_API_KEY") + if not api_key: + logger.warning("[vision.v2] DASHSCOPE_API_KEY 未配置，跳过 pro 兜底") return None + url = f"{_BASE_URL}/chat/completions" + payload: dict[str, Any] = { + "model": _PRO_MODEL, + "messages": [ + {"role": "system", "content": _PRO_SYSTEM}, + {"role": "user", "content": [ + {"type": "image_url", "image_url": {"url": img_url}}, + {"type": "text", "text": _PRO_USER}, + ]}, + ], + "temperature": 0.3, + "max_tokens": _DEFAULT_MAX_TOKENS, + "stream": False, + "enable_thinking": False, + } try: - template = get_template("image_analysis") - system = render_system_prompt(template) - user = render_user_prompt(template, image_count=1, industry="通用", image_urls=f"第1张：{img_url}") - except Exception as e: - logger.warning("[vision.vlm] 模板加载失败: %s", e) - return None - - client = get_doubao_client() - if not client.is_available: - return None - - use_model = model or DEFAULT_PRO_MODEL - _orig_retries = client.max_retries - client.max_retries = 0 - try: - raw = client.vision_completion( - messages=[{"role": "system", "content": system}, {"role": "user", "content": user}], - images=[img_url], - temperature=0.3, - max_tokens=DEFAULT_MAX_TOKENS, + r = httpx.post( + url, + headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"}, + json=payload, timeout=timeout, - model=use_model, ) - except Exception as e: - logger.warning("[vision.vlm] 图片 #%d pro VLM 调用失败 elapsed=%.1fs err=%s", idx, time.time() - t0, e) - client.max_retries = _orig_retries - return None - client.max_retries = _orig_retries - - elapsed = time.time() - t0 - if not raw: - logger.warning("[vision.vlm] 图片 #%d pro VLM 返回空 elapsed=%.1fs", idx, elapsed) - return None - - text = _strip_code_fence(raw) - l, r = text.find("{"), text.rfind("}") - if l >= 0 and r > l: - try: - obj = json.loads(text[l : r + 1]) - if isinstance(obj, dict): - logger.info("[vision.vlm] 图片 #%d pro VLM JSON 完成 elapsed=%.1fs", idx, elapsed) - return { - "name": obj.get("name") or "未识别", - "brand": obj.get("brand") or "无法判断", - "category": obj.get("category") or "无法判断", - "appearance": obj.get("appearance") or "无法判断", - "packaging": obj.get("packaging") or "无法判断", - "text_on_package": obj.get("text_on_package") or [], - "key_features": obj.get("key_features") or obj.get("features") or ["无法判断"], - "scene": obj.get("scene") or "通用", - "mood": obj.get("mood") or "", - "portrait_prompt": obj.get("portrait_prompt") or "无人像", - "summary": obj.get("summary") or f"{obj.get('brand','')} {obj.get('name','')}", - "_source": "vlm_pro_json", - } - except json.JSONDecodeError: - pass - - try: - result = _xml_to_product(text, idx) - result["_fallback_used"] = True - result["_pro_elapsed"] = round(elapsed, 2) + elapsed = time.time() - t0 + if r.status_code != 200: + logger.warning("[vision.v2] pro HTTP %d elapsed=%.1fs body=%s", r.status_code, elapsed, r.text[:200]) + return None + data = r.json() + raw = (data.get("choices") or [{}])[0].get("message", {}).get("content") + if not raw: + logger.warning("[vision.v2] pro 返回空 elapsed=%.1fs", elapsed) + return None + usage = data.get("usage") or {} logger.info( - "[vision.vlm] 图片 #%d pro VLM XML 完成 elapsed=%.2fs pp=%s", - idx, - elapsed, - (result.get("portrait_prompt") or "")[:40], + "[vision.v2] pro 完成 idx=%d model=%s elapsed=%.1fs in=%d out=%d", + idx, _PRO_MODEL, elapsed, + usage.get("prompt_tokens", 0), usage.get("completion_tokens", 0), ) - return result + text = _strip_code_fence(raw) + l, r_pos = text.find("{"), text.rfind("}") + if l < 0 or r_pos <= l: + logger.warning("[vision.v2] pro 无JSON elapsed=%.1fs head=%s", elapsed, raw[:200]) + return None + obj = json.loads(text[l:r_pos+1]) + if not isinstance(obj, dict): + return None + scene = obj.get("scene") or "通用" + mood = obj.get("mood") or "" + pp = _assemble_pp(obj) + has_person = obj.get("has_person", False) + has_product = obj.get("has_product", False) + name = obj.get("product_name") or "未识别" + brand = obj.get("brand") or "无法判断" + category = obj.get("category") or ("非产品图" if has_person and not has_product else "无法判断") + return { + "name": name, + "brand": brand, + "category": category, + "appearance": obj.get("outfit") or "无法判断", + "packaging": "无法判断", + "text_on_package": [], + "key_features": obj.get("key_features") or ["无法判断"], + "scene": scene, + "mood": mood, + "portrait_prompt": pp, + "summary": f"{brand} {name}" if name != "未识别" else "未识别", + "_source": "vlm_pro", + } except Exception as e: - logger.warning("[vision.vlm] 图片 #%d 解析失败 elapsed=%.1fs err=%s head=%s", idx, elapsed, e, raw[:200]) + logger.warning("[vision.v2] pro 异常 idx=%d elapsed=%.1fs err=%s", idx, time.time()-t0, e, exc_info=True) return None diff --git a/apps/worker/worker_app/tasks/vision/vlm_fast_json.py b/apps/worker/worker_app/tasks/vision/vlm_fast_json.py index 036ec8285..5fb245356 100644 --- a/apps/worker/worker_app/tasks/vision/vlm_fast_json.py +++ b/apps/worker/worker_app/tasks/vision/vlm_fast_json.py @@ -1,35 +1,43 @@ # -*- coding: utf-8 -*- -"""doubao-seed-2.1-lite 强约束 JSON-only 调用。 +"""V2 快速路径：qwen3.8-flash（阿里云百炼/DashScope）强约束 JSON-only 调用。 目标：替代"人体属性/商品检测/图像标签"三个火山不存在的专用云端 API。 设计要点： +- 直接用 httpx 发最小 payload 到 DashScope OpenAI 兼容 endpoint，不走 ai_client 包装 +- enable_thinking=false 关闭推理链（reasoning 是延迟主因） - system prompt 极致精简，只给字段 schema 和强约束（禁止自然语言、禁止 markdown） -- max_tokens=350（比旧 VLM 的 1200 小很多，降低延迟） -- temperature=0.1（极低，稳定输出 JSON） -- timeout=8s（够快，失败则由外层走 pro VLM 兜底） -- 期望返回纯 JSON object（无 ```json 包裹、无解释文字） +- max_tokens=350、temperature=0.1（稳定输出 JSON） +- timeout=8s（失败由外层走 pro 兜底） +- API Key 从环境变量 DASHSCOPE_API_KEY 读取 """ from __future__ import annotations import json import logging +import os import time from typing import Any logger = logging.getLogger(__name__) +# DashScope OpenAI 兼容 endpoint +_BASE_URL = "https://dashscope.aliyuncs.com/compatible-mode/v1" +_FAST_MODEL = "qwen3.8-flash" +_DEFAULT_TIMEOUT = 8 +_DEFAULT_MAX_TOKENS = 350 + # 极简 system prompt：只给字段定义 + 硬性输出要求 _FAST_SYSTEM = ( "你是图片结构化识别器。严格按下方 JSON schema 返回一个对象，不要任何解释、" "不要markdown、不要代码块、不要前后缀文字。字段值不确定时填 null 或空数组。\n" "{\n" - ' "has_person": true/false, // 图中是否有人\n' + ' "has_person": true/false,\n' ' "gender": "男"/"女"/null,\n' ' "age_range": "儿童"/"青少年"/"青年"/"中年"/"老年"/null,\n' ' "upper_wear": "上装款式，如T恤/衬衫/卫衣/毛衣/西装/夹克/连衣裙/吊带/背心/外套等",\n' ' "upper_color": "上装主色",\n' - ' "lower_wear": "下装款式，如牛仔裤/休闲裤/短裙/长裙/短裤/西裤/运动裤等；穿连衣裙时填null",\n' + ' "lower_wear": "下装款式；穿连衣裙时填null",\n' ' "lower_color": "下装主色",\n' ' "dress_color": "连衣裙主色（穿连衣裙时填）",\n' ' "accessories": ["眼镜"/"帽子"/"项链"/"耳环"/"背包"/"手表"等数组],\n' @@ -38,7 +46,7 @@ _FAST_SYSTEM = ( ' "pose": "姿势，如站立/坐姿/侧身/行走等",\n' ' "scene": "场景，如室内/街拍/户外/办公室/家居/海边/雪景/森林等",\n' ' "style": "风格，如休闲/商务/运动/复古/潮流/甜美/酷飒/优雅/街头/法式等",\n' - ' "has_product": true/false, // 是否有明确商品展示\n' + ' "has_product": true/false,\n' ' "category": "产品类目：服饰/鞋包/美妆/数码/食品/家居/配饰/母婴/非产品图",\n' ' "product_name": "产品名称，非产品图填null",\n' ' "brand": "品牌或文字标识，无则null",\n' @@ -51,21 +59,17 @@ _FAST_SYSTEM = ( _FAST_USER = "识别这张图片的人物穿搭与主体信息，只返回JSON对象。" -# 默认模型 -DEFAULT_LITE_MODEL = "doubao-seed-2-1-lite-260915" -DEFAULT_TIMEOUT = 8 -DEFAULT_MAX_TOKENS = 350 + +def _api_key() -> str | None: + return os.environ.get("DASHSCOPE_API_KEY") def _strip_code_fence(s: str) -> str: - """剥离 ```json ... ``` 包裹（即使要求纯 JSON，模型偶尔仍会包代码块）。""" s = s.strip() if s.startswith("```"): lines = s.split("\n") - # 去掉首行 ```json if lines and lines[0].startswith("```"): lines = lines[1:] - # 去掉尾行 ``` if lines and lines[-1].strip().startswith("```"): lines = lines[:-1] s = "\n".join(lines).strip() @@ -75,57 +79,37 @@ def _strip_code_fence(s: str) -> str: def call_fast_json( img_url: str, *, - model: str | None = None, - timeout: int = DEFAULT_TIMEOUT, - max_tokens: int = DEFAULT_MAX_TOKENS, + timeout: int = _DEFAULT_TIMEOUT, + max_tokens: int = _DEFAULT_MAX_TOKENS, ) -> dict[str, Any] | None: - """调用 lite VLM 返回结构化 dict；失败/非 JSON 返回 None。 - - 直接用 httpx 发最小 payload（关闭 thinking），不走 ai_client 包装： - - 关闭 thinking/推理链（reasoning_tokens 是延迟主因，单次要10-12s） - - 单次调用不重试（失败由外层走 pro 兜底） - - 温度=0.1 稳定输出 JSON - """ + """调用 qwen3.8-flash 返回结构化 dict；失败/非 JSON 返回 None。""" t0 = time.time() import httpx + api_key = _api_key() + if not api_key: + logger.warning("[vision.v2] DASHSCOPE_API_KEY 未配置，跳过 fast_json") + return None + + url = f"{_BASE_URL}/chat/completions" + payload: dict[str, Any] = { + "model": _FAST_MODEL, + "messages": [ + {"role": "system", "content": _FAST_SYSTEM}, + { + "role": "user", + "content": [ + {"type": "image_url", "image_url": {"url": img_url}}, + {"type": "text", "text": _FAST_USER}, + ], + }, + ], + "temperature": 0.1, + "max_tokens": max_tokens, + "stream": False, + "enable_thinking": False, + } try: - from packages.shared import get_shared_settings - - settings = get_shared_settings() - api_key = settings.doubao_api_key - base_url = (settings.doubao_base_url or "https://ark.cn-beijing.volces.com/api/v3").rstrip("/") - if not api_key: - logger.warning("[vision.v2] doubao api_key 未配置，跳过 fast_json") - return None - - use_model = model or DEFAULT_LITE_MODEL - url = f"{base_url}/chat/completions" - payload: dict[str, Any] = { - "model": use_model, - "messages": [ - {"role": "system", "content": _FAST_SYSTEM}, - { - "role": "user", - "content": [ - {"type": "image_url", "image_url": {"url": img_url}}, - {"type": "text", "text": _FAST_USER}, - ], - }, - ], - "temperature": 0.1, - "max_tokens": max_tokens, - "stream": False, - } - # 关键：关闭 thinking（避免产生 reasoning_tokens 拖慢响应） - # 方舟/豆包 2.x 模型支持 thinking.type=disabled - try: - payload["thinking"] = {"type": "disabled"} - except Exception: - pass - # 部分模型用 reasoning_effort 控制思考深度 - payload["reasoning_effort"] = "low" - resp = httpx.post( url, headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"}, @@ -133,71 +117,50 @@ def call_fast_json( timeout=timeout, ) elapsed = time.time() - t0 - if resp.status_code != 200: - logger.warning( - "[vision.v2] fast_json HTTP %d elapsed=%.1fs body=%s", resp.status_code, elapsed, resp.text[:200] + if resp.status_code == 400 and "enable_thinking" in resp.text[:300].lower(): + # 极少数 endpoint 版本不识别 enable_thinking，重试一次不带 + logger.warning("[vision.v2] fast_json HTTP 400 thinking 参数不兼容，重试 elapsed=%.1fs", elapsed) + payload.pop("enable_thinking", None) + resp = httpx.post( + url, + headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"}, + json=payload, + timeout=timeout, ) - # 如果400说明不支持thinking参数，降级重试一次 - if resp.status_code == 400 and "thinking" in resp.text.lower(): - payload.pop("thinking", None) - payload.pop("reasoning_effort", None) - time.time() - resp = httpx.post( - url, - headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"}, - json=payload, - timeout=timeout, - ) - elapsed = time.time() - t0 - if resp.status_code != 200: - logger.warning("[vision.v2] fast_json 降级后 HTTP %d elapsed=%.1fs", resp.status_code, elapsed) - return None - else: - return None + elapsed = time.time() - t0 + if resp.status_code != 200: + logger.warning("[vision.v2] fast_json HTTP %d elapsed=%.1fs body=%s", resp.status_code, elapsed, resp.text[:200]) + return None data = resp.json() raw = (data.get("choices") or [{}])[0].get("message", {}).get("content") - if raw is None: - logger.warning("[vision.v2] fast_json 返回 None elapsed=%.1fs", elapsed) + if not raw: + logger.warning("[vision.v2] fast_json 返回空 elapsed=%.1fs", elapsed) return None usage = data.get("usage") or {} + reasoning_tokens = usage.get("reasoning_tokens", 0) + ctd = usage.get("completion_tokens_details") or {} + if not reasoning_tokens: + reasoning_tokens = ctd.get("reasoning_tokens", 0) logger.info( - "[vision.v2] fast_json 直连完成 model=%s elapsed=%.1fs in=%d out=%d reasoning=%d", - use_model, - elapsed, - usage.get("prompt_tokens", 0), - usage.get("completion_tokens", 0), - usage.get("reasoning_tokens", 0), + "[vision.v2] fast_json 完成 model=%s elapsed=%.1fs in=%d out=%d reasoning=%d", + _FAST_MODEL, elapsed, + usage.get("prompt_tokens", 0), usage.get("completion_tokens", 0), reasoning_tokens, ) - elapsed = time.time() - t0 - if raw is None: - logger.warning("[vision.v2] fast_json 返回 None elapsed=%.1fs model=%s", elapsed, use_model) - return None - text = _strip_code_fence(raw) - # 截到第一个 { 和最后一个 } 之间，容忍前后偶发文字 - l = text.find("{") - r = text.rfind("}") + l, r = text.find("{"), text.rfind("}") if l >= 0 and r > l: text = text[l : r + 1] try: obj = json.loads(text) except json.JSONDecodeError: - logger.warning( - "[vision.v2] fast_json JSON 解析失败 elapsed=%.1fs head=%s", - elapsed, - raw[:200], - ) + logger.warning("[vision.v2] fast_json JSON 解析失败 elapsed=%.1fs head=%s", elapsed, raw[:200]) return None if not isinstance(obj, dict): logger.warning("[vision.v2] fast_json 非 dict: %s", type(obj)) return None logger.info( - "[vision.v2] fast_json 完成 model=%s elapsed=%.1fs has_person=%s has_product=%s category=%s", - use_model, - elapsed, - obj.get("has_person"), - obj.get("has_product"), - obj.get("category"), + "[vision.v2] fast_json 完成 elapsed=%.1fs has_person=%s has_product=%s category=%s", + elapsed, obj.get("has_person"), obj.get("has_product"), obj.get("category"), ) return obj except Exception as e: