mirror of
https://github.com/primedigitaltech/market-assistant.git
synced 2026-07-21 23:41:39 +08:00
refactor(pipeline): OpenAI 网关与配料视觉抽离至 openai_gateway
新增 pipeline/openai_gateway：credentials、text_chat、ingredients_op、ingredients_defaults 等；AI_crawler 仅加载 .env 与重导出。CrawlerOpenAiCompatible 与京东详情/流水线改走 import pipeline.openai_gateway，不再依赖 sys.path 注入爬虫目录。llm/providers/shared 转重导 gateway。详情脚本补充 backend 入 path。 Made-with: Cursor
This commit is contained in:
hub-gif
parent
33bf73e3ba
commit
9fd32a3f84
@ -2,43 +2,23 @@
|
||||
"""
|
||||
从本地图片路径或图片 URL 调用 OpenAI 兼容多模态接口，提取配料表等；并提供**纯文本** ``chat/completions`` 供报告/策略等场景复用。
|
||||
|
||||
**密钥与网关仅通过环境变量配置**（勿写入代码）：
|
||||
**实现**在 ``backend/pipeline/openai_gateway``；本文件负责：加载根目录 ``.env``、将 ``backend`` 加入 ``sys.path``、
|
||||
命令行试跑、以及**同名符号**重导（兼容历史 ``import AI_crawler``）。
|
||||
|
||||
- ``OPENAI_API_KEY``：API Key（必填）
|
||||
- ``OPENAI_BASE_URL``：网关根地址，如 ``https://llm.example.com/v1``（必填，勿尾斜杠多余路径）
|
||||
- ``OPENAI_VISION_MODEL``：视觉模型名（可选，默认 ``Qwen/Qwen3-Omni-30B-A3B``）
|
||||
环境变量说明见原仓库文档与 ``openai_gateway`` 模块注释。
|
||||
|
||||
**纯文本调用**（``chat_completion_text``）优先使用：
|
||||
|
||||
- ``OPENAI_TEXT_MODEL`` 或 ``LLM_TEXT_MODEL``；未设置时依次回退到 ``OPENAI_VISION_MODEL``、``LLM_MODEL``、上述默认。
|
||||
|
||||
兼容别名（二选一即可）：``LLM_API_KEY``、``LLM_BASE_URL``、``LLM_MODEL``。
|
||||
|
||||
- ``LLM_CONTEXT_WINDOW`` / ``OPENAI_CONTEXT_WINDOW``：模型上下文 token 上限（默认 ``32768``）。``chat_completion_text`` 会按正文长度**保守估算**输入 token，并把 ``max_tokens`` 收紧到不超过「上限 − 估算输入」，避免网关报 ``max_tokens … is too large``（如 LiteLLM ``ContextWindowExceededError``）。
|
||||
|
||||
上述变量与 Django 共用 **一份** ``market_assistant/.env``（与本脚本所在 ``backend`` 的上三级目录下的 ``.env``；需 ``pip install python-dotenv``）。
|
||||
|
||||
**运行方式**：在下方「运行配置」里改好 ``IMAGE_SOURCE`` 等变量后，直接执行 ``python AI_crawler.py``，无需命令行参数。
|
||||
**运行方式**：在下方改 ``IMAGE_SOURCE`` 后执行 ``python AI_crawler.py``，无需命令行参数。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import base64
|
||||
import os
|
||||
import re
|
||||
import sys
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
import requests
|
||||
|
||||
_SCRIPT_DIR = Path(__file__).resolve().parent
|
||||
# backend/crawler_copy/jd_pc_search -> parents[3] == market_assistant
|
||||
_MA_ROOT = Path(__file__).resolve().parents[3]
|
||||
|
||||
|
||||
def _load_market_assistant_dotenv() -> None:
|
||||
"""先于 LOW_GI_PROJECT_ROOT 解析加载 ``market_assistant/.env``（唯一配置源）。"""
|
||||
try:
|
||||
from dotenv import load_dotenv
|
||||
except ImportError:
|
||||
@ -49,740 +29,57 @@ def _load_market_assistant_dotenv() -> None:
|
||||
|
||||
|
||||
_load_market_assistant_dotenv()
|
||||
_BACKEND = Path(__file__).resolve().parents[2]
|
||||
if str(_BACKEND) not in sys.path:
|
||||
sys.path.insert(0, str(_BACKEND))
|
||||
|
||||
from _low_gi_root import low_gi_project_root # noqa: E402
|
||||
|
||||
_PROJECT_ROOT = low_gi_project_root()
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 运行配置（按需修改；启动时不要求命令行参数）
|
||||
# ---------------------------------------------------------------------------
|
||||
# 必填：本地图片路径，或 http(s) 图片链接（如京东主图 / 详情图）
|
||||
import requests # noqa: E402
|
||||
|
||||
from pipeline.openai_gateway import ( # noqa: E402
|
||||
REASON_NO_BODY_URLS,
|
||||
REASON_NO_VISION_API,
|
||||
chat_completion_text,
|
||||
extract_ingredients_from_body_image_urls_reversed,
|
||||
extract_ingredients_from_body_image_urls_reversed_with_source,
|
||||
extract_ingredients_from_image,
|
||||
normalize_ingredients_text_for_csv,
|
||||
normalize_message_content,
|
||||
parse_joined_image_urls,
|
||||
resolve_credentials,
|
||||
resolve_text_model_name,
|
||||
sanitize_vision_ingredients_output,
|
||||
strip_outer_markdown_fence,
|
||||
)
|
||||
from pipeline.openai_gateway.constants import ( # noqa: E402
|
||||
DEFAULT_MODEL,
|
||||
DEFAULT_USER_AGENT,
|
||||
)
|
||||
from pipeline.openai_gateway.credentials import _resolve_credentials # noqa: E402
|
||||
from pipeline.openai_gateway.ingredients_defaults import ( # noqa: E402
|
||||
IMAGE_REFERER,
|
||||
MAX_TOKENS,
|
||||
PROMPT_DEFAULT,
|
||||
QWEN_OMNI_TEMPLATE,
|
||||
TEMPERATURE,
|
||||
USER_PROMPT,
|
||||
)
|
||||
from pipeline.openai_gateway.ingredients_op import ( # noqa: E402
|
||||
_ingredient_extraction_acceptable,
|
||||
)
|
||||
|
||||
# 与早先脚本一致：供 ``python AI_crawler.py`` 单图试跑；多数字段与 ``ingredients_defaults`` 同义
|
||||
IMAGE_SOURCE = "https://img30.360buyimg.com/sku/jfs/t1/390444/8/13018/103574/6982e951Fc44d9d7b/00d62ee56189d75d.jpg.avif"
|
||||
# IMAGE_SOURCE = "https://img30.360buyimg.com/sku/jfs/t1/382894/31/7432/241977/694cf41aFa27be91e/00d63164ffeb8b46.jpg.avif"
|
||||
|
||||
# 提示词：留空则使用 ``PROMPT_DEFAULT``
|
||||
USER_PROMPT = ""
|
||||
PROMPT_DEFAULT = (
|
||||
"请识别图片中的配料表，只输出配料列表本身，不要将菜品做法、步骤、用料等认为是配料表，不要误识别为食谱；用中文逗号或顿号分隔，"
|
||||
"输出为连续一段文字，不要使用多行换行（避免与食谱、做法步骤混淆）。"
|
||||
"每种原料名称只出现一次，禁止重复罗列同一添加剂（如磷酸三钾、磷酸三钠等勿循环抄写多遍）；"
|
||||
"若图为表格中多行同类添加剂，可概括为「食品添加剂（按国家标准使用）」或合并为一句，勿展开成数百字重复。"
|
||||
"【禁止猜测】必须严格依据图中清晰可见的印刷文字归纳；不得根据商品品类、常识或模糊字迹推测、补全、编造任何原料。"
|
||||
"若本图无配料表、仅有产品信息/广告、文字被裁切、过小、模糊到无法逐字确认，或你只能「猜」出部分内容，则禁止输出配料列表："
|
||||
"请只输出且仅输出一句「无法识别图片中的配料表」（不要解释、不要道歉长文、不要列出疑似项）。"
|
||||
)
|
||||
|
||||
# 拉取远程图时的 Referer（京东图床一般需类似商城域名）
|
||||
IMAGE_REFERER = "https://www.jd.com/"
|
||||
|
||||
TEMPERATURE = 0.0
|
||||
MAX_TOKENS = 2048
|
||||
# 部分 Qwen 网关需要关闭 thinking
|
||||
QWEN_OMNI_TEMPLATE = False
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
DEFAULT_MODEL = "Qwen/Qwen3-Omni-30B-A3B"
|
||||
DEFAULT_USER_AGENT = (
|
||||
"Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 "
|
||||
"(KHTML, like Gecko) Chrome/120.0.0.0 Safari/537.36"
|
||||
)
|
||||
|
||||
|
||||
def _normalize_chat_content(content: Any) -> str:
|
||||
"""
|
||||
兼容 OpenAI 兼容网关：``message.content`` 可能是 str，也可能是
|
||||
``[{type:text, text:...}, ...]``；避免对 list 误用 ``.strip()`` 或得到怪异字符串。
|
||||
"""
|
||||
if content is None:
|
||||
return ""
|
||||
if isinstance(content, str):
|
||||
return content.strip()
|
||||
if isinstance(content, list):
|
||||
parts: list[str] = []
|
||||
for item in content:
|
||||
if isinstance(item, dict):
|
||||
if item.get("type") == "text":
|
||||
parts.append(str(item.get("text") or ""))
|
||||
elif "text" in item:
|
||||
parts.append(str(item.get("text") or ""))
|
||||
elif isinstance(item, str):
|
||||
parts.append(item)
|
||||
return "".join(parts).strip()
|
||||
return str(content).strip()
|
||||
|
||||
|
||||
def normalize_ingredients_text_for_csv(text: str) -> str:
|
||||
"""
|
||||
将配料 OCR 结果压成**单行**，便于 ``detail_ware_export.csv`` / 合并表展示。
|
||||
模型常按「一行一项」输出食谱或列表，会产生多换行；合并为非换行文本，行间用中文分号分隔。
|
||||
"""
|
||||
t = (text or "").replace("\r\n", "\n").replace("\r", "\n").strip()
|
||||
if not t:
|
||||
return ""
|
||||
lines = [ln.strip() for ln in t.split("\n") if ln.strip()]
|
||||
if len(lines) <= 1:
|
||||
return lines[0] if lines else ""
|
||||
return "；".join(lines)
|
||||
|
||||
|
||||
def _split_ingredient_segments(text: str) -> list[str]:
|
||||
"""按常见分隔符拆成原料小段（用于检测尾部循环复读）。"""
|
||||
t = (text or "").strip()
|
||||
if not t:
|
||||
return []
|
||||
return [p.strip() for p in re.split(r"[；、,，]+", t) if p.strip()]
|
||||
|
||||
|
||||
def sanitize_vision_ingredients_output(text: str) -> str:
|
||||
"""
|
||||
清洗多模态配料识别结果：去掉尾部引号、切除「仅两三种词循环数百次」的模型复读尾巴、超长截断。
|
||||
|
||||
典型故障：真实配料后无限重复「磷酸三钾、磷酸三钠…」，仍因前半段通过业务校验。
|
||||
"""
|
||||
t = (text or "").strip()
|
||||
_trail_q = frozenset({'"', "'", "\u201c", "\u201d", "\u2018", "\u2019", "\uff02"})
|
||||
while t and t[-1] in _trail_q:
|
||||
t = t[:-1].strip()
|
||||
|
||||
segs = _split_ingredient_segments(t)
|
||||
if not segs:
|
||||
return ""
|
||||
|
||||
min_spam_run = 28
|
||||
cut_i = len(segs)
|
||||
for i in range(0, max(0, len(segs) - min_spam_run + 1)):
|
||||
suf = segs[i:]
|
||||
if len(suf) >= min_spam_run and len(set(suf)) <= 3:
|
||||
cut_i = i
|
||||
break
|
||||
segs = segs[:cut_i]
|
||||
if not segs:
|
||||
return ""
|
||||
|
||||
t = "、".join(segs)
|
||||
|
||||
# 字符级兜底：同一短词组高频重复（未按顿号切分时）
|
||||
t = re.sub(
|
||||
r"(磷酸三[钾钠][、,，]?\s*){35,}",
|
||||
"磷酸三钾、磷酸三钠等（按国家标准使用）",
|
||||
t,
|
||||
)
|
||||
|
||||
max_chars = 3200
|
||||
if len(t) > max_chars:
|
||||
cut = t[:max_chars]
|
||||
last = max(cut.rfind("、"), cut.rfind("，"), cut.rfind(","), cut.rfind("；"))
|
||||
if last > max_chars // 2:
|
||||
t = cut[: last + 1] + "…（已截断）"
|
||||
else:
|
||||
t = cut + "…（已截断）"
|
||||
return t.strip()
|
||||
|
||||
|
||||
def _resolve_credentials(
|
||||
api_key: str | None,
|
||||
base_url: str | None,
|
||||
model: str | None,
|
||||
) -> tuple[str, str, str]:
|
||||
"""凭证只从环境变量（及可选函数参数）读取，不在代码中写死。"""
|
||||
key = (
|
||||
(api_key or "").strip()
|
||||
or (os.environ.get("OPENAI_API_KEY") or os.environ.get("LLM_API_KEY") or "").strip()
|
||||
)
|
||||
base = (
|
||||
(base_url or "").strip().rstrip("/")
|
||||
or (
|
||||
os.environ.get("OPENAI_BASE_URL") or os.environ.get("LLM_BASE_URL") or ""
|
||||
).strip().rstrip("/")
|
||||
)
|
||||
m = (
|
||||
(model or "").strip()
|
||||
or (
|
||||
os.environ.get("OPENAI_VISION_MODEL")
|
||||
or os.environ.get("LLM_MODEL")
|
||||
or DEFAULT_MODEL
|
||||
).strip()
|
||||
)
|
||||
if not key:
|
||||
raise ValueError("请设置环境变量 OPENAI_API_KEY（或 LLM_API_KEY）")
|
||||
if not base:
|
||||
raise ValueError(
|
||||
"请设置环境变量 OPENAI_BASE_URL（或 LLM_BASE_URL），例如 https://your-gateway.com/v1"
|
||||
)
|
||||
return key, base, m
|
||||
|
||||
|
||||
def resolve_text_model_name(model: str | None = None) -> str:
|
||||
"""
|
||||
文本补全所用模型：显式 ``model`` 优先，否则读环境变量（见模块文档）。
|
||||
"""
|
||||
m = (model or "").strip()
|
||||
if m:
|
||||
return m
|
||||
for env in (
|
||||
"OPENAI_TEXT_MODEL",
|
||||
"LLM_TEXT_MODEL",
|
||||
"OPENAI_VISION_MODEL",
|
||||
"LLM_MODEL",
|
||||
):
|
||||
v = (os.environ.get(env) or "").strip()
|
||||
if v:
|
||||
return v
|
||||
return DEFAULT_MODEL
|
||||
_normalize_chat_content = normalize_message_content
|
||||
|
||||
|
||||
def _estimate_chat_input_tokens(system_prompt: str, user_prompt: str) -> int:
|
||||
"""
|
||||
保守估算 system+user 的 token 数（无 tiktoken 时用于约束 max_tokens）。
|
||||
中文/JSON 往往高于「英文 4 字符/token」；系数偏大宁可少给 max_tokens，避免 400。
|
||||
"""
|
||||
total_chars = len(system_prompt or "") + len(user_prompt or "")
|
||||
return int(total_chars * 0.55) + 512
|
||||
from pipeline.openai_gateway.estimate import estimate_chat_input_tokens
|
||||
|
||||
|
||||
def strip_outer_markdown_fence(text: str) -> str:
|
||||
"""若模型用 ``` / ```markdown 包裹全文，去掉最外层围栏。"""
|
||||
t = (text or "").strip()
|
||||
if not t.startswith("```"):
|
||||
return t
|
||||
lines = t.split("\n")
|
||||
if lines and lines[0].strip().startswith("```"):
|
||||
lines = lines[1:]
|
||||
while lines and lines[-1].strip() == "```":
|
||||
lines = lines[:-1]
|
||||
return "\n".join(lines).strip()
|
||||
|
||||
|
||||
def chat_completion_text(
|
||||
*,
|
||||
system_prompt: str,
|
||||
user_prompt: str,
|
||||
api_key: str | None = None,
|
||||
base_url: str | None = None,
|
||||
model: str | None = None,
|
||||
temperature: float = 0.2,
|
||||
max_tokens: int = 8192,
|
||||
timeout: int = 300,
|
||||
extra_json: dict[str, Any] | None = None,
|
||||
) -> str:
|
||||
"""
|
||||
OpenAI 兼容网关的**纯文本**多轮占位为 system + user 各一条，与 ``extract_ingredients_from_image`` 共用凭证与端点。
|
||||
返回助手消息正文（已 ``strip`` / 兼容 list 型 content）。
|
||||
"""
|
||||
k, b, _ = _resolve_credentials(api_key, base_url, None)
|
||||
m = resolve_text_model_name(model)
|
||||
body: dict[str, Any] = {
|
||||
"model": m,
|
||||
"messages": [
|
||||
{"role": "system", "content": system_prompt},
|
||||
{"role": "user", "content": user_prompt},
|
||||
],
|
||||
"temperature": temperature,
|
||||
"max_tokens": max_tokens,
|
||||
}
|
||||
if extra_json:
|
||||
body.update(extra_json)
|
||||
# 避免 max_tokens 大于「上下文 − 输入」时网关 400（如 LiteLLM ContextWindowExceededError）
|
||||
ctx_raw = (
|
||||
os.environ.get("LLM_CONTEXT_WINDOW")
|
||||
or os.environ.get("OPENAI_CONTEXT_WINDOW")
|
||||
or "32768"
|
||||
).strip()
|
||||
try:
|
||||
context_window = max(4096, int(ctx_raw))
|
||||
except ValueError:
|
||||
context_window = 32768
|
||||
buf = 256
|
||||
input_est = _estimate_chat_input_tokens(system_prompt, user_prompt)
|
||||
if input_est >= context_window - buf - 256:
|
||||
raise ValueError(
|
||||
f"提示词过长（估算输入约 {input_est} tokens，上下文上限 {context_window}），"
|
||||
"请缩小报告/摘要输入或换更大上下文的模型；也可设置环境变量 LLM_CONTEXT_WINDOW。"
|
||||
)
|
||||
avail = context_window - input_est - buf
|
||||
want = int(body.get("max_tokens") or max_tokens)
|
||||
body["max_tokens"] = max(256, min(want, max(avail, 256)))
|
||||
r = requests.post(
|
||||
f"{b}/chat/completions",
|
||||
headers={
|
||||
"Authorization": f"Bearer {k}",
|
||||
"Content-Type": "application/json",
|
||||
},
|
||||
json=body,
|
||||
timeout=timeout,
|
||||
)
|
||||
try:
|
||||
r.raise_for_status()
|
||||
except requests.HTTPError as e:
|
||||
snippet = ""
|
||||
if e.response is not None:
|
||||
snippet = (e.response.text or "")[:1200].replace("\r\n", "\n").replace("\n", " ")
|
||||
if snippet:
|
||||
raise requests.HTTPError(
|
||||
f"{e!s} | body: {snippet}",
|
||||
response=e.response,
|
||||
request=e.request,
|
||||
) from e
|
||||
raise
|
||||
data = r.json()
|
||||
msg = (data.get("choices") or [{}])[0].get("message") or {}
|
||||
return _normalize_chat_content(msg.get("content"))
|
||||
|
||||
|
||||
def _mime_for_path(path: str) -> str:
|
||||
ext = path.lower().rsplit(".", 1)[-1]
|
||||
return {
|
||||
"jpg": "image/jpeg",
|
||||
"jpeg": "image/jpeg",
|
||||
"png": "image/png",
|
||||
"webp": "image/webp",
|
||||
"gif": "image/gif",
|
||||
"avif": "image/avif",
|
||||
}.get(ext, "image/jpeg")
|
||||
|
||||
|
||||
def _mime_from_response(url: str, content_type: str | None) -> str:
|
||||
if content_type and content_type.lower().startswith("image/"):
|
||||
return content_type.split(";")[0].strip().lower()
|
||||
u = url.lower().split("?")[0]
|
||||
for suf, mime in (
|
||||
(".png", "image/png"),
|
||||
(".webp", "image/webp"),
|
||||
(".avif", "image/avif"),
|
||||
(".gif", "image/gif"),
|
||||
(".jpg", "image/jpeg"),
|
||||
(".jpeg", "image/jpeg"),
|
||||
):
|
||||
if u.endswith(suf):
|
||||
return mime
|
||||
return "image/jpeg"
|
||||
|
||||
|
||||
def image_to_data_url(
|
||||
source: str,
|
||||
*,
|
||||
referer: str = "https://www.jd.com/",
|
||||
timeout: int = 60,
|
||||
) -> tuple[str, str]:
|
||||
"""
|
||||
``source`` 为本地路径或以 http(s) 开头的 URL。
|
||||
返回 (data_url, 来源说明)。
|
||||
"""
|
||||
s = source.strip()
|
||||
if s.lower().startswith(("http://", "https://")):
|
||||
headers = {
|
||||
"User-Agent": DEFAULT_USER_AGENT,
|
||||
"Accept": "image/avif,image/webp,image/*,*/*;q=0.8",
|
||||
"Referer": referer,
|
||||
}
|
||||
r = requests.get(s, headers=headers, timeout=timeout, allow_redirects=True)
|
||||
r.raise_for_status()
|
||||
mime = _mime_from_response(s, r.headers.get("Content-Type"))
|
||||
b64 = base64.standard_b64encode(r.content).decode("ascii")
|
||||
return f"data:{mime};base64,{b64}", f"url:{s[:80]}"
|
||||
|
||||
with open(s, "rb") as f:
|
||||
raw = f.read()
|
||||
mime = _mime_for_path(s)
|
||||
b64 = base64.standard_b64encode(raw).decode("ascii")
|
||||
return f"data:{mime};base64,{b64}", f"file:{s}"
|
||||
|
||||
|
||||
def extract_ingredients_from_image(
|
||||
image_path_or_url: str,
|
||||
*,
|
||||
api_key: str | None = None,
|
||||
base_url: str | None = None,
|
||||
model: str | None = None,
|
||||
user_prompt: str | None = None,
|
||||
temperature: float = 0.0,
|
||||
max_tokens: int = 2048,
|
||||
referer: str = "https://www.jd.com/",
|
||||
extra_json: dict[str, Any] | None = None,
|
||||
prompt_default: str | None = None,
|
||||
) -> str:
|
||||
"""
|
||||
从本地图片路径或图片 URL 识别配料表（可改 ``user_prompt`` 扩展为营养成分表等）。
|
||||
未传入 ``api_key`` / ``base_url`` / ``model`` 时从环境变量读取。
|
||||
返回值为经 ``normalize_ingredients_text_for_csv`` 处理后的**单行**文本，便于写入 CSV。
|
||||
"""
|
||||
k, b, m = _resolve_credentials(api_key, base_url, model)
|
||||
|
||||
data_url, _src = image_to_data_url(image_path_or_url, referer=referer)
|
||||
|
||||
_fallback = (
|
||||
prompt_default
|
||||
or "请识别图片中的配料表，只输出配料列表，不要误识别为做法用料；用逗号或顿号分隔为一段，不要换行分段。"
|
||||
)
|
||||
prompt = user_prompt if user_prompt is not None and str(user_prompt).strip() else _fallback
|
||||
|
||||
body: dict[str, Any] = {
|
||||
"model": m,
|
||||
"messages": [
|
||||
{
|
||||
"role": "user",
|
||||
"content": [
|
||||
{"type": "image_url", "image_url": {"url": data_url}},
|
||||
{"type": "text", "text": prompt},
|
||||
],
|
||||
}
|
||||
],
|
||||
"temperature": temperature,
|
||||
"max_tokens": max_tokens,
|
||||
}
|
||||
if extra_json:
|
||||
body.update(extra_json)
|
||||
|
||||
r = requests.post(
|
||||
f"{b}/chat/completions",
|
||||
headers={
|
||||
"Authorization": f"Bearer {k}",
|
||||
"Content-Type": "application/json",
|
||||
},
|
||||
json=body,
|
||||
timeout=120,
|
||||
)
|
||||
r.raise_for_status()
|
||||
data = r.json()
|
||||
msg = (data.get("choices") or [{}])[0].get("message") or {}
|
||||
raw = normalize_ingredients_text_for_csv(_normalize_chat_content(msg.get("content")))
|
||||
return sanitize_vision_ingredients_output(raw)
|
||||
|
||||
|
||||
def parse_joined_image_urls(joined: str) -> list[str]:
|
||||
"""
|
||||
解析详情 DOM 拼出的 URL 串（与列 ``detail_body_ingredients`` 在「仅 URL」阶段同形：分号、换行分隔的 http(s) 链接）。
|
||||
保持从前到后的顺序；去重不在这里做（上游已去重）。
|
||||
"""
|
||||
t = (joined or "").strip()
|
||||
if not t:
|
||||
return []
|
||||
t = t.replace("\r\n", "\n").replace("\r", "\n")
|
||||
parts = re.split(r"\s*;\s*|\s*\n\s*", t)
|
||||
out: list[str] = []
|
||||
for p in parts:
|
||||
u = p.strip()
|
||||
if u.startswith(("http://", "https://")):
|
||||
out.append(u)
|
||||
return out
|
||||
|
||||
|
||||
def _looks_like_recipe_or_dish_prep(text: str) -> bool:
|
||||
"""
|
||||
判断模型输出是否更像**菜谱/做法备料**（详情图里常见），而非包装「配料表」。
|
||||
命中则不应写入 ``detail_body_ingredients``，继续尝试其它长图。
|
||||
"""
|
||||
t = (text or "").strip()
|
||||
if not t:
|
||||
return False
|
||||
|
||||
recipe_kw = (
|
||||
"做法",
|
||||
"制作步骤",
|
||||
"烹饪步骤",
|
||||
"第一步",
|
||||
"第二步",
|
||||
"第三步",
|
||||
"教程",
|
||||
"准备食材",
|
||||
"食材准备",
|
||||
"下锅",
|
||||
"翻炒",
|
||||
"煮熟",
|
||||
"大火烧开",
|
||||
"转小火",
|
||||
"装盘",
|
||||
"小贴士",
|
||||
"腌制",
|
||||
"爆香",
|
||||
"焯水",
|
||||
"切丝",
|
||||
"切丁",
|
||||
"切片",
|
||||
"打匀",
|
||||
"搅拌均匀",
|
||||
"油热",
|
||||
"调味",
|
||||
)
|
||||
if any(k in t for k in recipe_kw):
|
||||
return True
|
||||
|
||||
# 「葱花蒜末 各1勺」类菜谱用量
|
||||
if re.search(r"各[一二两三四五六七八九十\d零]+勺", t):
|
||||
return True
|
||||
|
||||
# 多条「短名称 + 数量 + 料理常用单位」并列（典型备料清单）
|
||||
dish_qty = re.findall(
|
||||
r"[^\n；。,，、]{1,14}\s+\d+(?:\.\d+)?\s*[个只根块片勺条袋包杯碗适量克gG毫升mlML]{1,4}",
|
||||
t,
|
||||
)
|
||||
if len(dish_qty) >= 2:
|
||||
return True
|
||||
|
||||
# 半块/半根等家常分量词 + 生鲜食材名（包装配料表极少这样写）
|
||||
if "半块" in t and re.search(r"鸡胸|鸡腿|牛肉|猪肉|黄瓜|番茄|土豆|豆腐", t):
|
||||
return True
|
||||
if "半根" in t and re.search(r"黄瓜|胡萝卜|玉米|香肠|葱", t):
|
||||
return True
|
||||
|
||||
# 规范化后的「A；B；C；…」若多段都很短且多段含数字，多为做法用料枚举
|
||||
parts = [p.strip() for p in t.split("；") if p.strip()]
|
||||
if len(parts) >= 4:
|
||||
short_with_digit = [p for p in parts if len(p) <= 24 and re.search(r"\d", p)]
|
||||
if len(short_with_digit) >= 4:
|
||||
return True
|
||||
|
||||
# 多行/多段里至少 3 条「短句 + 数字 + 个根块勺克」
|
||||
lines = [ln.strip() for ln in re.split(r"[\n；]", t) if ln.strip()]
|
||||
if len(lines) >= 3:
|
||||
n_short_qty = sum(
|
||||
1
|
||||
for ln in lines
|
||||
if len(ln) <= 22
|
||||
and re.search(r"\d", ln)
|
||||
and re.search(r"[个只根块片勺克gG]", ln)
|
||||
)
|
||||
if n_short_qty >= 3:
|
||||
return True
|
||||
|
||||
return False
|
||||
|
||||
|
||||
def _looks_like_packaged_ingredient_enumeration(text: str) -> bool:
|
||||
"""
|
||||
视觉模型常把包装图上的「配料表」整段压成**逗号/顿号分隔的原料枚举**，丢掉标题与含量行。
|
||||
此类文本与菜谱备料（鸡胸、黄瓜、葱花等）可区分时，视为有效配料信号。
|
||||
"""
|
||||
t = (text or "").strip()
|
||||
if not t:
|
||||
return False
|
||||
|
||||
parts = [p.strip() for p in re.split(r"[,，、;；]", t) if p.strip()]
|
||||
if len(parts) < 3:
|
||||
return False
|
||||
|
||||
# 多段像「家常备料」则不走此路（避免鸡胸、鸡蛋、黄瓜…误过）
|
||||
recipe_seg = re.compile(
|
||||
r"鸡胸|鸡腿|牛腩|牛肉|五花肉|里脊|鸡蛋|鸭蛋|皮蛋|黄瓜|番茄|西红柿|土豆|马铃薯|"
|
||||
r"葱花|蒜末|姜丝|小米椒|青椒|洋葱|胡萝卜|生菜|菠菜|白菜|芹菜|香菜|小葱|"
|
||||
r"面条$|挂面|粉条|粉丝"
|
||||
)
|
||||
n_recipe_like = sum(1 for p in parts if recipe_seg.search(p))
|
||||
if n_recipe_like >= 2:
|
||||
return False
|
||||
|
||||
# 工业化配料常见子串（粉体、纤维、添加剂类别、粮谷原料等）
|
||||
industrial = re.compile(
|
||||
r"食用|食品添加|麦麸|纤维|淀粉|魔芋|提取物|谷朊|谷胱|麸皮|糖浆|山梨|麦芽|柠檬酸|碳酸|"
|
||||
r"酵母|乳粉|全脂|脱脂|果胶|黄原|卡拉胶|海藻酸|小麦|面粉|荞麦|燕麦|藜麦|青稞|糙米|黑米|"
|
||||
r"棕榈|植物油|精炼油|氢化|起酥|可可脂"
|
||||
)
|
||||
n_industrial = sum(1 for p in parts if industrial.search(p))
|
||||
if len(parts) >= 4 and n_industrial >= 2:
|
||||
return True
|
||||
if len(parts) >= 3 and n_industrial >= 3:
|
||||
return True
|
||||
return False
|
||||
|
||||
|
||||
def _has_packaged_ingredient_table_signals(text: str) -> bool:
|
||||
"""
|
||||
正向判断：是否像**包装配料表**——标题+含量、行内含量、或（多段工业化原料枚举）。
|
||||
|
||||
仅 OCR 出一段家常食材名、无上述结构时，返回 False。
|
||||
"""
|
||||
t = (text or "").strip()
|
||||
if not t:
|
||||
return False
|
||||
|
||||
# 行内「××（含量≥50%）」等，常见于包装，不强制出现「配料表」标题
|
||||
if re.search(
|
||||
r"[\u4e00-\u9fff\w·．\d]{1,18}[（(]\s*含量\s*[≥>=＝]?\s*[\d.]+\s*%?\s*[)）]",
|
||||
t,
|
||||
):
|
||||
return True
|
||||
|
||||
label = bool(
|
||||
re.search(r"配料表", t)
|
||||
or re.search(r"配\s*料\s*[:：]", t)
|
||||
or re.search(r"原\s*料\s*[:：]", t)
|
||||
or re.search(r"食品添加剂", t)
|
||||
or re.search(r"产品\s*配\s*料", t)
|
||||
)
|
||||
|
||||
# 「含量」相关信息：百分比、不等式、法规用语、添加量表述等（不含单独「50克」类菜谱用量）
|
||||
content = bool(
|
||||
re.search(r"含量", t)
|
||||
or re.search(r"添加量", t)
|
||||
or re.search(r"\d+(?:\.\d+)?\s*[％%]", t)
|
||||
or re.search(r"[≥＞>]\s*[\d.]+", t)
|
||||
or re.search(r"按\s*添\s*加\s*量\s*递\s*减", t)
|
||||
)
|
||||
|
||||
if label and content:
|
||||
return True
|
||||
|
||||
# 模型只输出「原料1,原料2,…」时仍可能是正规配料表
|
||||
if _looks_like_packaged_ingredient_enumeration(t):
|
||||
return True
|
||||
|
||||
return False
|
||||
|
||||
|
||||
def _ingredient_extraction_acceptable(text: str) -> bool:
|
||||
"""粗判模型输出是否像有效配料信息（过滤拒识句、伪列表、过短碎片、菜谱备料）。
|
||||
|
||||
通过条件之一：配料表标题+含量类信号；行内「××（含量≥x%）」；或多段工业化原料枚举（见
|
||||
``_looks_like_packaged_ingredient_enumeration``，用于模型只输出逗号分隔原料、丢掉标题时）。
|
||||
"""
|
||||
t = (text or "").strip()
|
||||
if len(t) < 6:
|
||||
return False
|
||||
# 模型偶发输出类似 Python 列表的字符串，或 JSON 数组形态
|
||||
if re.match(r"^\s*\[.*\]\s*$", t):
|
||||
return False
|
||||
refuse = (
|
||||
"无法识别",
|
||||
"没有配料",
|
||||
"看不清",
|
||||
"不存在配料",
|
||||
"未在图中",
|
||||
"未在图片",
|
||||
"抱歉，我",
|
||||
"抱歉，无法",
|
||||
"不能识别",
|
||||
"没有识别到",
|
||||
"图中没有",
|
||||
"图片中没有",
|
||||
"无配料",
|
||||
"未见配料",
|
||||
)
|
||||
if any(x in t for x in refuse):
|
||||
return False
|
||||
# 真配料表通常含分隔符或足够长；避免「无」「暂无」等被当成命中
|
||||
if t in ("无", "暂无", "没有", "无。", "无，"):
|
||||
return False
|
||||
if _looks_like_recipe_or_dish_prep(t):
|
||||
return False
|
||||
if not _has_packaged_ingredient_table_signals(t):
|
||||
return False
|
||||
tail = _split_ingredient_segments(t)
|
||||
if len(tail) >= 32:
|
||||
if len(set(tail[-32:])) <= 3:
|
||||
return False
|
||||
sep_chars = "，、,;；"
|
||||
if len(t) < 18 and not any(c in t for c in sep_chars):
|
||||
return False
|
||||
return True
|
||||
|
||||
|
||||
REASON_NO_BODY_URLS = "【未识别到配料】未解析到任何详情长图 URL。"
|
||||
REASON_NO_VISION_API = (
|
||||
"【未识别到配料】未配置多模态 API（需环境变量 OPENAI_API_KEY + OPENAI_BASE_URL，"
|
||||
"或 LLM_API_KEY + LLM_BASE_URL）。"
|
||||
)
|
||||
|
||||
|
||||
def extract_ingredients_from_body_image_urls_reversed_with_source(
|
||||
urls_joined: str,
|
||||
*,
|
||||
referer: str | None = None,
|
||||
user_prompt: str | None = None,
|
||||
prompt_default: str | None = None,
|
||||
temperature: float | None = None,
|
||||
max_tokens: int | None = None,
|
||||
extra_json: dict[str, Any] | None = None,
|
||||
) -> tuple[str, str | None]:
|
||||
"""
|
||||
与 ``extract_ingredients_from_body_image_urls_reversed`` 相同逻辑；额外返回命中配料时所用的**图片 URL**
|
||||
（自后向前首次通过校验的那张）。未命中或失败时第二项为 ``None``。
|
||||
"""
|
||||
urls = parse_joined_image_urls(urls_joined)
|
||||
if not urls:
|
||||
return REASON_NO_BODY_URLS, None
|
||||
try:
|
||||
_resolve_credentials(None, None, None)
|
||||
except ValueError:
|
||||
return REASON_NO_VISION_API, None
|
||||
|
||||
ref = (referer if referer is not None else IMAGE_REFERER) or "https://www.jd.com/"
|
||||
temp = float(temperature) if temperature is not None else float(TEMPERATURE)
|
||||
mt = int(max_tokens) if max_tokens is not None else int(MAX_TOKENS)
|
||||
extra = extra_json
|
||||
if extra is None and QWEN_OMNI_TEMPLATE:
|
||||
extra = {"chat_template_kwargs": {"enable_thinking": False}}
|
||||
|
||||
pu = user_prompt if user_prompt is not None else ((USER_PROMPT or "").strip() or None)
|
||||
pd = prompt_default if prompt_default is not None else PROMPT_DEFAULT
|
||||
|
||||
n = len(urls)
|
||||
n_err = 0
|
||||
n_rejected = 0
|
||||
for url in reversed(urls):
|
||||
try:
|
||||
text = extract_ingredients_from_image(
|
||||
url,
|
||||
user_prompt=pu,
|
||||
referer=ref.strip(),
|
||||
temperature=temp,
|
||||
max_tokens=mt,
|
||||
extra_json=extra,
|
||||
prompt_default=pd,
|
||||
)
|
||||
except Exception:
|
||||
n_err += 1
|
||||
continue
|
||||
t = (text or "").strip()
|
||||
if _ingredient_extraction_acceptable(t):
|
||||
return t, url
|
||||
if t:
|
||||
n_rejected += 1
|
||||
|
||||
parts = [
|
||||
f"【未识别到配料】已对 {n} 张详情长图自后向前依次尝试（命中即停），未得到有效配料表。"
|
||||
]
|
||||
if n_err:
|
||||
parts.append(f" 请求异常 {n_err} 次。")
|
||||
if n_rejected:
|
||||
parts.append(f" 有 {n_rejected} 次返回未通过配料校验。")
|
||||
if not n_err and not n_rejected:
|
||||
parts.append(" 模型返回均为空或过短。")
|
||||
return "".join(parts), None
|
||||
|
||||
|
||||
def extract_ingredients_from_body_image_urls_reversed(
|
||||
urls_joined: str,
|
||||
*,
|
||||
referer: str | None = None,
|
||||
user_prompt: str | None = None,
|
||||
prompt_default: str | None = None,
|
||||
temperature: float | None = None,
|
||||
max_tokens: int | None = None,
|
||||
extra_json: dict[str, Any] | None = None,
|
||||
) -> str:
|
||||
"""
|
||||
对 URL 串拆出的链接 **从后往前**依次调用视觉模型：**首次**通过校验的配料文本立即返回（省时间）。
|
||||
|
||||
若始终无命中：返回以 ``【未识别到配料】`` 开头的原因说明（**不再返回空串**）。
|
||||
未配置 API 时返回 ``REASON_NO_VISION_API``。
|
||||
|
||||
命中条件（见 ``_ingredient_extraction_acceptable``）：须像**包装配料表**——「配料/含量」标题结构、
|
||||
``××（含量≥x%）``、或多段工业化原料逗号/顿号枚举（模型常省略标题）；纯家常备料（鸡胸、黄瓜、葱花等）
|
||||
仍丢弃并试下一张图。
|
||||
|
||||
若需同时得到所用图片 URL，请用 ``extract_ingredients_from_body_image_urls_reversed_with_source``。
|
||||
"""
|
||||
text, _ = extract_ingredients_from_body_image_urls_reversed_with_source(
|
||||
urls_joined,
|
||||
referer=referer,
|
||||
user_prompt=user_prompt,
|
||||
prompt_default=prompt_default,
|
||||
temperature=temperature,
|
||||
max_tokens=max_tokens,
|
||||
extra_json=extra_json,
|
||||
)
|
||||
return text
|
||||
return estimate_chat_input_tokens(system_prompt, user_prompt)
|
||||
|
||||
|
||||
def main() -> None:
|
||||
@ -803,7 +100,7 @@ def main() -> None:
|
||||
sys.exit(2)
|
||||
|
||||
prompt_use = (USER_PROMPT or "").strip() or None
|
||||
extra = None
|
||||
extra: dict[str, Any] | None = None
|
||||
if QWEN_OMNI_TEMPLATE:
|
||||
extra = {"chat_template_kwargs": {"enable_thinking": False}}
|
||||
|
||||
|
||||
@ -21,7 +21,7 @@
|
||||
- 若 ``OUTPUT_SKU_AND_BODY_IMAGES_ONLY=False``：**原始接口 JSON** 单 SKU ``OUT``、批量 ``OUT_DIR`` / ``ware_{sku}.json``；解析扁平含全部 ``detail_*``；汇总表 ``OUT_PARSED_CSV``。
|
||||
|
||||
**解析 API**：``flatten_ware_business`` / ``parse_ware_business_response_text`` / ``ware_parsed_row``，
|
||||
列 ``detail_body_ingredients`` 为配料表文本（由 ``#detail-main`` 长图经 ``AI_crawler`` 自后向前多模态识别）；列 ``detail_body_ingredients_source_url`` 为**实际用于识别**的那张长图 URL（命中即停）。未配置 API 或识别失败时配料列为原因说明、图源列为空。内部 ``meta["detail_body_image_urls"]`` 仍为全部长图 URL 串，仅供解析用。
|
||||
列 ``detail_body_ingredients`` 为配料表文本（由 ``#detail-main`` 长图经 ``pipeline.openai_gateway`` 自后向前多模态识别）；列 ``detail_body_ingredients_source_url`` 为**实际用于识别**的那张长图 URL（命中即停）。未配置 API 或识别失败时配料列为原因说明、图源列为空。内部 ``meta["detail_body_image_urls"]`` 仍为全部长图 URL 串，仅供解析用。
|
||||
|
||||
Cookie：``../common/jd_cookie.txt``（或配置项 ``COOKIE_FILE`` / ``COOKIE_OVERRIDE``），经 ``add_cookies`` 注入。
|
||||
|
||||
@ -51,6 +51,10 @@ from playwright.sync_api import sync_playwright
|
||||
_JD_PC_SEARCH = Path(__file__).resolve().parents[1]
|
||||
if str(_JD_PC_SEARCH) not in sys.path:
|
||||
sys.path.insert(0, str(_JD_PC_SEARCH))
|
||||
# ``import pipeline.openai_gateway`` 需将 backend 根目录加入 path（Django 启动时已有）
|
||||
_backend_root = Path(__file__).resolve().parents[3]
|
||||
if str(_backend_root) not in sys.path:
|
||||
sys.path.insert(0, str(_backend_root))
|
||||
from _low_gi_root import low_gi_project_root # noqa: E402
|
||||
|
||||
_PROJECT_ROOT = low_gi_project_root()
|
||||
@ -92,7 +96,7 @@ LOG_API_M_JD_TRACE = False
|
||||
# COLLECT_DETAIL_MAIN_IMAGE_URLS：True 时在点击商品详情 tab 并等待后，从 #detail-main 抽取图文 URL（style 中 background-image、img src、zbViewWeChatMiniImages），
|
||||
# 补全为 https 后写入 meta（见 DETAIL_BODY_IMAGE_URL_SEPARATOR）；再经多模态写入列 detail_body_ingredients（配料文本）与 detail_body_ingredients_source_url（命中图源）
|
||||
COLLECT_DETAIL_MAIN_IMAGE_URLS = True
|
||||
# EXTRACT_INGREDIENTS_FROM_DETAIL_BODY_IMAGES：True 时 detail_body_ingredients 为配料表文本、detail_body_ingredients_source_url 为命中图源；False 时配料列恒为空、图源列恒为空。需 .env 中 OPENAI_* / LLM_*（见上级目录 AI_crawler.py）
|
||||
# EXTRACT_INGREDIENTS_FROM_DETAIL_BODY_IMAGES：True 时 detail_body_ingredients 为配料表文本、detail_body_ingredients_source_url 为命中图源；False 时配料列恒为空、图源列恒为空。需 .env 中 OPENAI_* / LLM_*（见 pipeline.openai_gateway）
|
||||
EXTRACT_INGREDIENTS_FROM_DETAIL_BODY_IMAGES = True
|
||||
# DETAIL_BODY_IMAGE_URL_SEPARATOR：写入 CSV/JSON 单单元格时的分隔符
|
||||
DETAIL_BODY_IMAGE_URL_SEPARATOR = "; "
|
||||
@ -348,9 +352,9 @@ def main() -> None:
|
||||
vision_ok = False
|
||||
if EXTRACT_INGREDIENTS_FROM_DETAIL_BODY_IMAGES:
|
||||
try:
|
||||
import AI_crawler as vision_mod # noqa: WPS433
|
||||
import pipeline.openai_gateway as vision_mod # noqa: WPS433
|
||||
|
||||
vision_mod._resolve_credentials(None, None, None)
|
||||
vision_mod.resolve_credentials(None, None, None)
|
||||
vision_ok = True
|
||||
except Exception as e:
|
||||
print(
|
||||
|
||||
@ -318,9 +318,9 @@ def main(keyword: str | None = None) -> Path:
|
||||
_ingredient_vision_ok = False
|
||||
if EXTRACT_INGREDIENTS_FROM_DETAIL_BODY_IMAGES:
|
||||
try:
|
||||
import AI_crawler as _ac_mod # noqa: WPS433
|
||||
import pipeline.openai_gateway as _ac_mod # noqa: WPS433
|
||||
|
||||
_ac_mod._resolve_credentials(None, None, None)
|
||||
_ac_mod.resolve_credentials(None, None, None)
|
||||
_ingredient_vision_ok = True
|
||||
except Exception as e:
|
||||
print(
|
||||
|
||||
@ -1,6 +1,6 @@
|
||||
"""
|
||||
竞品报告 / 策略稿的**大模型生成**：通过 ``crawler_copy/jd_pc_search/AI_crawler`` 的
|
||||
``chat_completion_text`` 调用，与配料识别共用网关与密钥。
|
||||
竞品报告 / 策略稿的**大模型生成**：经 ``pipeline.llm`` → ``openai_gateway.chat_completion_text``（OpenAI 兼容），
|
||||
与配料多模态识别共用 `OPENAI_*` / `LLM_*` 环境配置。
|
||||
|
||||
实现已拆分为子模块（``llm_client``、``generate_*``），本模块保留对外符号以兼容
|
||||
``from pipeline.llm.generate import …`` 与测试中的 patch 路径。
|
||||
|
||||
@ -1,11 +1,10 @@
|
||||
"""具体大模型通道实现：经统一协议暴露给 `factory`。"""
|
||||
from __future__ import annotations
|
||||
|
||||
from .crawler_openai_compatible import CrawlerOpenAiCompatibleTextLlm, ensure_ai_crawler_path
|
||||
from .crawler_openai_compatible import CrawlerOpenAiCompatibleTextLlm
|
||||
from .openai_official_chatgpt import OpenAiOfficialChatGptTextLlm
|
||||
|
||||
__all__ = [
|
||||
"CrawlerOpenAiCompatibleTextLlm",
|
||||
"OpenAiOfficialChatGptTextLlm",
|
||||
"ensure_ai_crawler_path",
|
||||
]
|
||||
|
||||
@ -1,26 +1,15 @@
|
||||
"""
|
||||
经 `crawler_copy/jd_pc_search/AI_crawler.chat_completion_text` 访问 OpenAI 兼容网关（与配料识别等共用凭据与配置）。
|
||||
经 ``pipeline.openai_gateway.text_chat.chat_completion_text`` 访问 OpenAI 兼容网关（与配料识别等共用环境变量，不再 import 爬虫目录）。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import os
|
||||
import sys
|
||||
from pathlib import Path
|
||||
|
||||
from django.conf import settings
|
||||
from pipeline.openai_gateway import chat_completion_text
|
||||
|
||||
from ..shared.token_heuristics import estimate_crawler_style_input_tokens
|
||||
|
||||
|
||||
def ensure_ai_crawler_path() -> None:
|
||||
root = Path(settings.CRAWLER_JD_ROOT).resolve()
|
||||
if not root.is_dir():
|
||||
raise FileNotFoundError(f"爬虫副本目录不存在: {root}")
|
||||
rs = str(root)
|
||||
if rs not in sys.path:
|
||||
sys.path.insert(0, rs)
|
||||
|
||||
|
||||
def _llm_context_window_size_from_env() -> int:
|
||||
raw = (
|
||||
os.environ.get("LLM_CONTEXT_WINDOW")
|
||||
@ -35,8 +24,7 @@ def _llm_context_window_size_from_env() -> int:
|
||||
|
||||
class CrawlerOpenAiCompatibleTextLlm:
|
||||
"""
|
||||
文本任务默认后端：复用 `AI_crawler` 的 `chat_completion_text` 与上下文预检逻辑。
|
||||
更换为其它云厂商时，应新增独立适配器并在 `factory` 中注册，而非修改本类。
|
||||
文本任务默认后端：与历史 ``AI_crawler.chat_completion_text`` 行为一致，实现位于 ``pipeline.openai_gateway``。
|
||||
"""
|
||||
|
||||
def complete_text(
|
||||
@ -46,16 +34,13 @@ class CrawlerOpenAiCompatibleTextLlm:
|
||||
*,
|
||||
temperature: float | None = None,
|
||||
) -> str:
|
||||
ensure_ai_crawler_path()
|
||||
import AI_crawler as ac # noqa: WPS433
|
||||
|
||||
kwargs: dict[str, object] = {
|
||||
"system_prompt": system_prompt,
|
||||
"user_prompt": user_prompt,
|
||||
}
|
||||
if temperature is not None:
|
||||
kwargs["temperature"] = float(temperature)
|
||||
return ac.chat_completion_text(**kwargs)
|
||||
return chat_completion_text(**kwargs)
|
||||
|
||||
def estimate_input_tokens(self, system_prompt: str, user_prompt: str) -> int:
|
||||
return estimate_crawler_style_input_tokens(system_prompt, user_prompt)
|
||||
|
||||
@ -11,8 +11,10 @@ from typing import Any
|
||||
|
||||
import requests
|
||||
|
||||
from ..shared.openai_message_content import normalize_message_content
|
||||
from ..shared.token_heuristics import estimate_crawler_style_input_tokens
|
||||
from pipeline.openai_gateway.chat_content import normalize_message_content
|
||||
from pipeline.openai_gateway.estimate import (
|
||||
estimate_chat_input_tokens as estimate_crawler_style_input_tokens,
|
||||
)
|
||||
|
||||
_DEFAULT_BASE = "https://api.openai.com/v1"
|
||||
_DEFAULT_MODEL = "gpt-4o-mini"
|
||||
|
||||
@ -1,26 +1,8 @@
|
||||
"""
|
||||
解析 OpenAI chat.completions 返回的 `message.content`：可能是 str 或 part 列表。
|
||||
与 `AI_crawler._normalize_chat_content` 行为一致。
|
||||
解析 OpenAI `message.content`；实现位于 ``pipeline.openai_gateway.chat_content``，仅重导以保持与旧 import 路径兼容。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
from typing import Any
|
||||
from pipeline.openai_gateway.chat_content import normalize_message_content
|
||||
|
||||
|
||||
def normalize_message_content(content: Any) -> str:
|
||||
if content is None:
|
||||
return ""
|
||||
if isinstance(content, str):
|
||||
return content.strip()
|
||||
if isinstance(content, list):
|
||||
parts: list[str] = []
|
||||
for item in content:
|
||||
if isinstance(item, dict):
|
||||
if item.get("type") == "text":
|
||||
parts.append(str(item.get("text") or ""))
|
||||
elif "text" in item:
|
||||
parts.append(str(item.get("text") or ""))
|
||||
elif isinstance(item, str):
|
||||
parts.append(item)
|
||||
return "".join(parts).strip()
|
||||
return str(content).strip()
|
||||
__all__ = ["normalize_message_content"]
|
||||
|
||||
@ -1,15 +1,6 @@
|
||||
"""对模型原始输出做与业务无关的轻量归一化（不修改提示词）。"""
|
||||
"""对模型输出去围栏等；与 ``openai_gateway.text_chat.strip_outer_markdown_fence`` 同义。"""
|
||||
from __future__ import annotations
|
||||
|
||||
from pipeline.openai_gateway.text_chat import strip_outer_markdown_fence
|
||||
|
||||
def strip_outer_markdown_fence(text: str) -> str:
|
||||
"""若模型用 ``` / ```markdown 包裹全文，去掉最外层围栏。与 `AI_crawler.strip_outer_markdown_fence` 行为一致。"""
|
||||
t = (text or "").strip()
|
||||
if not t.startswith("```"):
|
||||
return t
|
||||
lines = t.split("\n")
|
||||
if lines and lines[0].strip().startswith("```"):
|
||||
lines = lines[1:]
|
||||
while lines and lines[-1].strip() == "```":
|
||||
lines = lines[:-1]
|
||||
return "\n".join(lines).strip()
|
||||
__all__ = ["strip_outer_markdown_fence"]
|
||||
|
||||
@ -1,10 +1,8 @@
|
||||
"""
|
||||
与 `crawler_copy/.../AI_crawler` 中 `_estimate_chat_input_tokens` 同口径的保守估算，
|
||||
供预检、策略档位与各适配器共用（无 tiktoken 时避免 max_tokens 400）。
|
||||
与 ``pipeline.openai_gateway.estimate.estimate_chat_input_tokens`` 同口径的保守估算。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
from pipeline.openai_gateway.estimate import estimate_chat_input_tokens as estimate_crawler_style_input_tokens
|
||||
|
||||
def estimate_crawler_style_input_tokens(system_prompt: str, user_prompt: str) -> int:
|
||||
total_chars = len(system_prompt or "") + len(user_prompt or "")
|
||||
return int(total_chars * 0.55) + 512
|
||||
__all__ = ["estimate_crawler_style_input_tokens"]
|
||||
|
||||
42
backend/pipeline/openai_gateway/__init__.py
Normal file
42
backend/pipeline/openai_gateway/__init__.py
Normal file
@ -0,0 +1,42 @@
|
||||
"""
|
||||
OpenAI 兼容网关（`chat/completions`）：纯文本、多模态配料、详情长图逆序等。
|
||||
|
||||
与 ``crawler_copy/.../AI_crawler`` 解耦，逻辑唯一来源在 ``pipeline.openai_gateway``。
|
||||
脚本 ``AI_crawler.py`` 仅作加载 .env 与命令行试跑、并向后兼容重导出名。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
from .chat_content import normalize_message_content
|
||||
from .credentials import (
|
||||
_resolve_credentials,
|
||||
resolve_credentials,
|
||||
resolve_text_model_name,
|
||||
)
|
||||
from .ingredients_op import (
|
||||
REASON_NO_BODY_URLS,
|
||||
REASON_NO_VISION_API,
|
||||
extract_ingredients_from_body_image_urls_reversed,
|
||||
extract_ingredients_from_body_image_urls_reversed_with_source,
|
||||
extract_ingredients_from_image,
|
||||
normalize_ingredients_text_for_csv,
|
||||
parse_joined_image_urls,
|
||||
sanitize_vision_ingredients_output,
|
||||
)
|
||||
from .text_chat import chat_completion_text, strip_outer_markdown_fence
|
||||
|
||||
__all__ = [
|
||||
"REASON_NO_BODY_URLS",
|
||||
"REASON_NO_VISION_API",
|
||||
"_resolve_credentials",
|
||||
"chat_completion_text",
|
||||
"extract_ingredients_from_body_image_urls_reversed",
|
||||
"extract_ingredients_from_body_image_urls_reversed_with_source",
|
||||
"extract_ingredients_from_image",
|
||||
"normalize_ingredients_text_for_csv",
|
||||
"normalize_message_content",
|
||||
"parse_joined_image_urls",
|
||||
"resolve_credentials",
|
||||
"resolve_text_model_name",
|
||||
"sanitize_vision_ingredients_output",
|
||||
"strip_outer_markdown_fence",
|
||||
]
|
||||
23
backend/pipeline/openai_gateway/chat_content.py
Normal file
23
backend/pipeline/openai_gateway/chat_content.py
Normal file
@ -0,0 +1,23 @@
|
||||
"""解析 `chat/completions` 返回的 `message.content`（str 或多段）。"""
|
||||
from __future__ import annotations
|
||||
|
||||
from typing import Any
|
||||
|
||||
|
||||
def normalize_message_content(content: Any) -> str:
|
||||
if content is None:
|
||||
return ""
|
||||
if isinstance(content, str):
|
||||
return content.strip()
|
||||
if isinstance(content, list):
|
||||
parts: list[str] = []
|
||||
for item in content:
|
||||
if isinstance(item, dict):
|
||||
if item.get("type") == "text":
|
||||
parts.append(str(item.get("text") or ""))
|
||||
elif "text" in item:
|
||||
parts.append(str(item.get("text") or ""))
|
||||
elif isinstance(item, str):
|
||||
parts.append(item)
|
||||
return "".join(parts).strip()
|
||||
return str(content).strip()
|
||||
8
backend/pipeline/openai_gateway/constants.py
Normal file
8
backend/pipeline/openai_gateway/constants.py
Normal file
@ -0,0 +1,8 @@
|
||||
"""OpenAI 兼容调用的缺省常数（可仍由环境变量覆盖）。"""
|
||||
from __future__ import annotations
|
||||
|
||||
DEFAULT_MODEL = "Qwen/Qwen3-Omni-30B-A3B"
|
||||
DEFAULT_USER_AGENT = (
|
||||
"Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 "
|
||||
"(KHTML, like Gecko) Chrome/120.0.0.0 Safari/537.36"
|
||||
)
|
||||
63
backend/pipeline/openai_gateway/credentials.py
Normal file
63
backend/pipeline/openai_gateway/credentials.py
Normal file
@ -0,0 +1,63 @@
|
||||
"""OpenAI 兼容网关：从环境或参数解析 API 凭证与模型名（文本 / 多模态共用）。"""
|
||||
from __future__ import annotations
|
||||
|
||||
import os
|
||||
|
||||
from .constants import DEFAULT_MODEL
|
||||
|
||||
|
||||
def _resolve_credentials(
|
||||
api_key: str | None,
|
||||
base_url: str | None,
|
||||
model: str | None,
|
||||
) -> tuple[str, str, str]:
|
||||
key = (
|
||||
(api_key or "").strip()
|
||||
or (os.environ.get("OPENAI_API_KEY") or os.environ.get("LLM_API_KEY") or "").strip()
|
||||
)
|
||||
base = (
|
||||
(base_url or "").strip().rstrip("/")
|
||||
or (
|
||||
os.environ.get("OPENAI_BASE_URL") or os.environ.get("LLM_BASE_URL") or ""
|
||||
).strip().rstrip("/")
|
||||
)
|
||||
m = (
|
||||
(model or "").strip()
|
||||
or (
|
||||
os.environ.get("OPENAI_VISION_MODEL")
|
||||
or os.environ.get("LLM_MODEL")
|
||||
or DEFAULT_MODEL
|
||||
).strip()
|
||||
)
|
||||
if not key:
|
||||
raise ValueError("请设置环境变量 OPENAI_API_KEY（或 LLM_API_KEY）")
|
||||
if not base:
|
||||
raise ValueError(
|
||||
"请设置环境变量 OPENAI_BASE_URL（或 LLM_BASE_URL），例如 https://your-gateway.com/v1"
|
||||
)
|
||||
return key, base, m
|
||||
|
||||
|
||||
def resolve_text_model_name(model: str | None = None) -> str:
|
||||
m = (model or "").strip()
|
||||
if m:
|
||||
return m
|
||||
for env in (
|
||||
"OPENAI_TEXT_MODEL",
|
||||
"LLM_TEXT_MODEL",
|
||||
"OPENAI_VISION_MODEL",
|
||||
"LLM_MODEL",
|
||||
):
|
||||
v = (os.environ.get(env) or "").strip()
|
||||
if v:
|
||||
return v
|
||||
return DEFAULT_MODEL
|
||||
|
||||
|
||||
def resolve_credentials(
|
||||
api_key: str | None = None,
|
||||
base_url: str | None = None,
|
||||
model: str | None = None,
|
||||
) -> tuple[str, str, str]:
|
||||
"""与历史 ``AI_crawler._resolve_credentials`` 行为一致，供业务显式预检（如多模态是否可用）。"""
|
||||
return _resolve_credentials(api_key, base_url, model)
|
||||
7
backend/pipeline/openai_gateway/estimate.py
Normal file
7
backend/pipeline/openai_gateway/estimate.py
Normal file
@ -0,0 +1,7 @@
|
||||
"""输入 token 保守估算（与历史 AI_crawler 同口径）。"""
|
||||
from __future__ import annotations
|
||||
|
||||
|
||||
def estimate_chat_input_tokens(system_prompt: str, user_prompt: str) -> int:
|
||||
total_chars = len(system_prompt or "") + len(user_prompt or "")
|
||||
return int(total_chars * 0.55) + 512
|
||||
20
backend/pipeline/openai_gateway/ingredients_defaults.py
Normal file
20
backend/pipeline/openai_gateway/ingredients_defaults.py
Normal file
@ -0,0 +1,20 @@
|
||||
"""详情页配料多模态识别：提示词、温度与 Referer 等（与 `AI_crawler` 顶部运行配置同口径）。"""
|
||||
from __future__ import annotations
|
||||
|
||||
# 与 AI_crawler 中 PROMPT_DEFAULT 保持同步（脚本入口仍可在本地覆盖）
|
||||
IMAGE_REFERER = "https://www.jd.com/"
|
||||
|
||||
USER_PROMPT = ""
|
||||
PROMPT_DEFAULT = (
|
||||
"请识别图片中的配料表，只输出配料列表本身，不要将菜品做法、步骤、用料等认为是配料表，不要误识别为食谱；用中文逗号或顿号分隔，"
|
||||
"输出为连续一段文字，不要使用多行换行（避免与食谱、做法步骤混淆）。"
|
||||
"每种原料名称只出现一次，禁止重复罗列同一添加剂（如磷酸三钾、磷酸三钠等勿循环抄写多遍）；"
|
||||
"若图为表格中多行同类添加剂，可概括为「食品添加剂（按国家标准使用）」或合并为一句，勿展开成数百字重复。"
|
||||
"【禁止猜测】必须严格依据图中清晰可见的印刷文字归纳；不得根据商品品类、常识或模糊字迹推测、补全、编造任何原料。"
|
||||
"若本图无配料表、仅有产品信息/广告、文字被裁切、过小、模糊到无法逐字确认，或你只能「猜」出部分内容，则禁止输出配料列表："
|
||||
"请只输出且仅输出一句「无法识别图片中的配料表」（不要解释、不要道歉长文、不要列出疑似项）。"
|
||||
)
|
||||
TEMPERATURE = 0.0
|
||||
MAX_TOKENS = 2048
|
||||
# 部分 Qwen 网关需要关闭 thinking
|
||||
QWEN_OMNI_TEMPLATE = False
|
||||
536
backend/pipeline/openai_gateway/ingredients_op.py
Normal file
536
backend/pipeline/openai_gateway/ingredients_op.py
Normal file
@ -0,0 +1,536 @@
|
||||
"""配料识别与详情长图逆序等逻辑。由 `AI_crawler` 抽离至 `pipeline.openai_gateway`。"""
|
||||
from __future__ import annotations
|
||||
|
||||
import base64
|
||||
import re
|
||||
from typing import Any
|
||||
|
||||
import requests
|
||||
|
||||
from . import ingredients_defaults as _d
|
||||
from .chat_content import normalize_message_content as _normalize_chat_content
|
||||
from .constants import DEFAULT_USER_AGENT
|
||||
from .credentials import _resolve_credentials
|
||||
|
||||
|
||||
def normalize_ingredients_text_for_csv(text: str) -> str:
|
||||
"""
|
||||
将配料 OCR 结果压成**单行**，便于 ``detail_ware_export.csv`` / 合并表展示。
|
||||
模型常按「一行一项」输出食谱或列表，会产生多换行；合并为非换行文本，行间用中文分号分隔。
|
||||
"""
|
||||
t = (text or "").replace("\r\n", "\n").replace("\r", "\n").strip()
|
||||
if not t:
|
||||
return ""
|
||||
lines = [ln.strip() for ln in t.split("\n") if ln.strip()]
|
||||
if len(lines) <= 1:
|
||||
return lines[0] if lines else ""
|
||||
return "；".join(lines)
|
||||
|
||||
|
||||
def _split_ingredient_segments(text: str) -> list[str]:
|
||||
"""按常见分隔符拆成原料小段（用于检测尾部循环复读）。"""
|
||||
t = (text or "").strip()
|
||||
if not t:
|
||||
return []
|
||||
return [p.strip() for p in re.split(r"[；、,，]+", t) if p.strip()]
|
||||
|
||||
|
||||
def sanitize_vision_ingredients_output(text: str) -> str:
|
||||
"""
|
||||
清洗多模态配料识别结果：去掉尾部引号、切除「仅两三种词循环数百次」的模型复读尾巴、超长截断。
|
||||
|
||||
典型故障：真实配料后无限重复「磷酸三钾、磷酸三钠…」，仍因前半段通过业务校验。
|
||||
"""
|
||||
t = (text or "").strip()
|
||||
_trail_q = frozenset({'"', "'", "\u201c", "\u201d", "\u2018", "\u2019", "\uff02"})
|
||||
while t and t[-1] in _trail_q:
|
||||
t = t[:-1].strip()
|
||||
|
||||
segs = _split_ingredient_segments(t)
|
||||
if not segs:
|
||||
return ""
|
||||
|
||||
min_spam_run = 28
|
||||
cut_i = len(segs)
|
||||
for i in range(0, max(0, len(segs) - min_spam_run + 1)):
|
||||
suf = segs[i:]
|
||||
if len(suf) >= min_spam_run and len(set(suf)) <= 3:
|
||||
cut_i = i
|
||||
break
|
||||
segs = segs[:cut_i]
|
||||
if not segs:
|
||||
return ""
|
||||
|
||||
t = "、".join(segs)
|
||||
|
||||
# 字符级兜底：同一短词组高频重复（未按顿号切分时）
|
||||
t = re.sub(
|
||||
r"(磷酸三[钾钠][、,，]?\s*){35,}",
|
||||
"磷酸三钾、磷酸三钠等（按国家标准使用）",
|
||||
t,
|
||||
)
|
||||
|
||||
max_chars = 3200
|
||||
if len(t) > max_chars:
|
||||
cut = t[:max_chars]
|
||||
last = max(cut.rfind("、"), cut.rfind("，"), cut.rfind(","), cut.rfind("；"))
|
||||
if last > max_chars // 2:
|
||||
t = cut[: last + 1] + "…（已截断）"
|
||||
else:
|
||||
t = cut + "…（已截断）"
|
||||
return t.strip()
|
||||
|
||||
|
||||
def _mime_for_path(path: str) -> str:
|
||||
ext = path.lower().rsplit(".", 1)[-1]
|
||||
return {
|
||||
"jpg": "image/jpeg",
|
||||
"jpeg": "image/jpeg",
|
||||
"png": "image/png",
|
||||
"webp": "image/webp",
|
||||
"gif": "image/gif",
|
||||
"avif": "image/avif",
|
||||
}.get(ext, "image/jpeg")
|
||||
|
||||
|
||||
def _mime_from_response(url: str, content_type: str | None) -> str:
|
||||
if content_type and content_type.lower().startswith("image/"):
|
||||
return content_type.split(";")[0].strip().lower()
|
||||
u = url.lower().split("?")[0]
|
||||
for suf, mime in (
|
||||
(".png", "image/png"),
|
||||
(".webp", "image/webp"),
|
||||
(".avif", "image/avif"),
|
||||
(".gif", "image/gif"),
|
||||
(".jpg", "image/jpeg"),
|
||||
(".jpeg", "image/jpeg"),
|
||||
):
|
||||
if u.endswith(suf):
|
||||
return mime
|
||||
return "image/jpeg"
|
||||
|
||||
|
||||
def image_to_data_url(
|
||||
source: str,
|
||||
*,
|
||||
referer: str = "https://www.jd.com/",
|
||||
timeout: int = 60,
|
||||
) -> tuple[str, str]:
|
||||
"""
|
||||
``source`` 为本地路径或以 http(s) 开头的 URL。
|
||||
返回 (data_url, 来源说明)。
|
||||
"""
|
||||
s = source.strip()
|
||||
if s.lower().startswith(("http://", "https://")):
|
||||
headers = {
|
||||
"User-Agent": DEFAULT_USER_AGENT,
|
||||
"Accept": "image/avif,image/webp,image/*,*/*;q=0.8",
|
||||
"Referer": referer,
|
||||
}
|
||||
r = requests.get(s, headers=headers, timeout=timeout, allow_redirects=True)
|
||||
r.raise_for_status()
|
||||
mime = _mime_from_response(s, r.headers.get("Content-Type"))
|
||||
b64 = base64.standard_b64encode(r.content).decode("ascii")
|
||||
return f"data:{mime};base64,{b64}", f"url:{s[:80]}"
|
||||
|
||||
with open(s, "rb") as f:
|
||||
raw = f.read()
|
||||
mime = _mime_for_path(s)
|
||||
b64 = base64.standard_b64encode(raw).decode("ascii")
|
||||
return f"data:{mime};base64,{b64}", f"file:{s}"
|
||||
|
||||
|
||||
def extract_ingredients_from_image(
|
||||
image_path_or_url: str,
|
||||
*,
|
||||
api_key: str | None = None,
|
||||
base_url: str | None = None,
|
||||
model: str | None = None,
|
||||
user_prompt: str | None = None,
|
||||
temperature: float = 0.0,
|
||||
max_tokens: int = 2048,
|
||||
referer: str = "https://www.jd.com/",
|
||||
extra_json: dict[str, Any] | None = None,
|
||||
prompt_default: str | None = None,
|
||||
) -> str:
|
||||
"""
|
||||
从本地图片路径或图片 URL 识别配料表（可改 ``user_prompt`` 扩展为营养成分表等）。
|
||||
未传入 ``api_key`` / ``base_url`` / ``model`` 时从环境变量读取。
|
||||
返回值为经 ``normalize_ingredients_text_for_csv`` 处理后的**单行**文本，便于写入 CSV。
|
||||
"""
|
||||
k, b, m = _resolve_credentials(api_key, base_url, model)
|
||||
|
||||
data_url, _src = image_to_data_url(image_path_or_url, referer=referer)
|
||||
|
||||
_fallback = (
|
||||
prompt_default
|
||||
or "请识别图片中的配料表，只输出配料列表，不要误识别为做法用料；用逗号或顿号分隔为一段，不要换行分段。"
|
||||
)
|
||||
prompt = user_prompt if user_prompt is not None and str(user_prompt).strip() else _fallback
|
||||
|
||||
body: dict[str, Any] = {
|
||||
"model": m,
|
||||
"messages": [
|
||||
{
|
||||
"role": "user",
|
||||
"content": [
|
||||
{"type": "image_url", "image_url": {"url": data_url}},
|
||||
{"type": "text", "text": prompt},
|
||||
],
|
||||
}
|
||||
],
|
||||
"temperature": temperature,
|
||||
"max_tokens": max_tokens,
|
||||
}
|
||||
if extra_json:
|
||||
body.update(extra_json)
|
||||
|
||||
r = requests.post(
|
||||
f"{b}/chat/completions",
|
||||
headers={
|
||||
"Authorization": f"Bearer {k}",
|
||||
"Content-Type": "application/json",
|
||||
},
|
||||
json=body,
|
||||
timeout=120,
|
||||
)
|
||||
r.raise_for_status()
|
||||
data = r.json()
|
||||
msg = (data.get("choices") or [{}])[0].get("message") or {}
|
||||
raw = normalize_ingredients_text_for_csv(_normalize_chat_content(msg.get("content")))
|
||||
return sanitize_vision_ingredients_output(raw)
|
||||
|
||||
|
||||
def parse_joined_image_urls(joined: str) -> list[str]:
|
||||
"""
|
||||
解析详情 DOM 拼出的 URL 串（与列 ``detail_body_ingredients`` 在「仅 URL」阶段同形：分号、换行分隔的 http(s) 链接）。
|
||||
保持从前到后的顺序；去重不在这里做（上游已去重）。
|
||||
"""
|
||||
t = (joined or "").strip()
|
||||
if not t:
|
||||
return []
|
||||
t = t.replace("\r\n", "\n").replace("\r", "\n")
|
||||
parts = re.split(r"\s*;\s*|\s*\n\s*", t)
|
||||
out: list[str] = []
|
||||
for p in parts:
|
||||
u = p.strip()
|
||||
if u.startswith(("http://", "https://")):
|
||||
out.append(u)
|
||||
return out
|
||||
|
||||
|
||||
def _looks_like_recipe_or_dish_prep(text: str) -> bool:
|
||||
"""
|
||||
判断模型输出是否更像**菜谱/做法备料**（详情图里常见），而非包装「配料表」。
|
||||
命中则不应写入 ``detail_body_ingredients``，继续尝试其它长图。
|
||||
"""
|
||||
t = (text or "").strip()
|
||||
if not t:
|
||||
return False
|
||||
|
||||
recipe_kw = (
|
||||
"做法",
|
||||
"制作步骤",
|
||||
"烹饪步骤",
|
||||
"第一步",
|
||||
"第二步",
|
||||
"第三步",
|
||||
"教程",
|
||||
"准备食材",
|
||||
"食材准备",
|
||||
"下锅",
|
||||
"翻炒",
|
||||
"煮熟",
|
||||
"大火烧开",
|
||||
"转小火",
|
||||
"装盘",
|
||||
"小贴士",
|
||||
"腌制",
|
||||
"爆香",
|
||||
"焯水",
|
||||
"切丝",
|
||||
"切丁",
|
||||
"切片",
|
||||
"打匀",
|
||||
"搅拌均匀",
|
||||
"油热",
|
||||
"调味",
|
||||
)
|
||||
if any(k in t for k in recipe_kw):
|
||||
return True
|
||||
|
||||
# 「葱花蒜末 各1勺」类菜谱用量
|
||||
if re.search(r"各[一二两三四五六七八九十\d零]+勺", t):
|
||||
return True
|
||||
|
||||
# 多条「短名称 + 数量 + 料理常用单位」并列（典型备料清单）
|
||||
dish_qty = re.findall(
|
||||
r"[^\n；。,，、]{1,14}\s+\d+(?:\.\d+)?\s*[个只根块片勺条袋包杯碗适量克gG毫升mlML]{1,4}",
|
||||
t,
|
||||
)
|
||||
if len(dish_qty) >= 2:
|
||||
return True
|
||||
|
||||
# 半块/半根等家常分量词 + 生鲜食材名（包装配料表极少这样写）
|
||||
if "半块" in t and re.search(r"鸡胸|鸡腿|牛肉|猪肉|黄瓜|番茄|土豆|豆腐", t):
|
||||
return True
|
||||
if "半根" in t and re.search(r"黄瓜|胡萝卜|玉米|香肠|葱", t):
|
||||
return True
|
||||
|
||||
# 规范化后的「A；B；C；…」若多段都很短且多段含数字，多为做法用料枚举
|
||||
parts = [p.strip() for p in t.split("；") if p.strip()]
|
||||
if len(parts) >= 4:
|
||||
short_with_digit = [p for p in parts if len(p) <= 24 and re.search(r"\d", p)]
|
||||
if len(short_with_digit) >= 4:
|
||||
return True
|
||||
|
||||
# 多行/多段里至少 3 条「短句 + 数字 + 个根块勺克」
|
||||
lines = [ln.strip() for ln in re.split(r"[\n；]", t) if ln.strip()]
|
||||
if len(lines) >= 3:
|
||||
n_short_qty = sum(
|
||||
1
|
||||
for ln in lines
|
||||
if len(ln) <= 22
|
||||
and re.search(r"\d", ln)
|
||||
and re.search(r"[个只根块片勺克gG]", ln)
|
||||
)
|
||||
if n_short_qty >= 3:
|
||||
return True
|
||||
|
||||
return False
|
||||
|
||||
|
||||
def _looks_like_packaged_ingredient_enumeration(text: str) -> bool:
|
||||
"""
|
||||
视觉模型常把包装图上的「配料表」整段压成**逗号/顿号分隔的原料枚举**，丢掉标题与含量行。
|
||||
此类文本与菜谱备料（鸡胸、黄瓜、葱花等）可区分时，视为有效配料信号。
|
||||
"""
|
||||
t = (text or "").strip()
|
||||
if not t:
|
||||
return False
|
||||
|
||||
parts = [p.strip() for p in re.split(r"[,，、;；]", t) if p.strip()]
|
||||
if len(parts) < 3:
|
||||
return False
|
||||
|
||||
# 多段像「家常备料」则不走此路（避免鸡胸、鸡蛋、黄瓜…误过）
|
||||
recipe_seg = re.compile(
|
||||
r"鸡胸|鸡腿|牛腩|牛肉|五花肉|里脊|鸡蛋|鸭蛋|皮蛋|黄瓜|番茄|西红柿|土豆|马铃薯|"
|
||||
r"葱花|蒜末|姜丝|小米椒|青椒|洋葱|胡萝卜|生菜|菠菜|白菜|芹菜|香菜|小葱|"
|
||||
r"面条$|挂面|粉条|粉丝"
|
||||
)
|
||||
n_recipe_like = sum(1 for p in parts if recipe_seg.search(p))
|
||||
if n_recipe_like >= 2:
|
||||
return False
|
||||
|
||||
# 工业化配料常见子串（粉体、纤维、添加剂类别、粮谷原料等）
|
||||
industrial = re.compile(
|
||||
r"食用|食品添加|麦麸|纤维|淀粉|魔芋|提取物|谷朊|谷胱|麸皮|糖浆|山梨|麦芽|柠檬酸|碳酸|"
|
||||
r"酵母|乳粉|全脂|脱脂|果胶|黄原|卡拉胶|海藻酸|小麦|面粉|荞麦|燕麦|藜麦|青稞|糙米|黑米|"
|
||||
r"棕榈|植物油|精炼油|氢化|起酥|可可脂"
|
||||
)
|
||||
n_industrial = sum(1 for p in parts if industrial.search(p))
|
||||
if len(parts) >= 4 and n_industrial >= 2:
|
||||
return True
|
||||
if len(parts) >= 3 and n_industrial >= 3:
|
||||
return True
|
||||
return False
|
||||
|
||||
|
||||
def _has_packaged_ingredient_table_signals(text: str) -> bool:
|
||||
"""
|
||||
正向判断：是否像**包装配料表**——标题+含量、行内含量、或（多段工业化原料枚举）。
|
||||
|
||||
仅 OCR 出一段家常食材名、无上述结构时，返回 False。
|
||||
"""
|
||||
t = (text or "").strip()
|
||||
if not t:
|
||||
return False
|
||||
|
||||
# 行内「××（含量≥50%）」等，常见于包装，不强制出现「配料表」标题
|
||||
if re.search(
|
||||
r"[\u4e00-\u9fff\w·．\d]{1,18}[（(]\s*含量\s*[≥>=＝]?\s*[\d.]+\s*%?\s*[)）]",
|
||||
t,
|
||||
):
|
||||
return True
|
||||
|
||||
label = bool(
|
||||
re.search(r"配料表", t)
|
||||
or re.search(r"配\s*料\s*[:：]", t)
|
||||
or re.search(r"原\s*料\s*[:：]", t)
|
||||
or re.search(r"食品添加剂", t)
|
||||
or re.search(r"产品\s*配\s*料", t)
|
||||
)
|
||||
|
||||
# 「含量」相关信息：百分比、不等式、法规用语、添加量表述等（不含单独「50克」类菜谱用量）
|
||||
content = bool(
|
||||
re.search(r"含量", t)
|
||||
or re.search(r"添加量", t)
|
||||
or re.search(r"\d+(?:\.\d+)?\s*[％%]", t)
|
||||
or re.search(r"[≥＞>]\s*[\d.]+", t)
|
||||
or re.search(r"按\s*添\s*加\s*量\s*递\s*减", t)
|
||||
)
|
||||
|
||||
if label and content:
|
||||
return True
|
||||
|
||||
# 模型只输出「原料1,原料2,…」时仍可能是正规配料表
|
||||
if _looks_like_packaged_ingredient_enumeration(t):
|
||||
return True
|
||||
|
||||
return False
|
||||
|
||||
|
||||
def _ingredient_extraction_acceptable(text: str) -> bool:
|
||||
"""粗判模型输出是否像有效配料信息（过滤拒识句、伪列表、过短碎片、菜谱备料）。
|
||||
|
||||
通过条件之一：配料表标题+含量类信号；行内「××（含量≥x%）」；或多段工业化原料枚举（见
|
||||
``_looks_like_packaged_ingredient_enumeration``，用于模型只输出逗号分隔原料、丢掉标题时）。
|
||||
"""
|
||||
t = (text or "").strip()
|
||||
if len(t) < 6:
|
||||
return False
|
||||
# 模型偶发输出类似 Python 列表的字符串，或 JSON 数组形态
|
||||
if re.match(r"^\s*\[.*\]\s*$", t):
|
||||
return False
|
||||
refuse = (
|
||||
"无法识别",
|
||||
"没有配料",
|
||||
"看不清",
|
||||
"不存在配料",
|
||||
"未在图中",
|
||||
"未在图片",
|
||||
"抱歉，我",
|
||||
"抱歉，无法",
|
||||
"不能识别",
|
||||
"没有识别到",
|
||||
"图中没有",
|
||||
"图片中没有",
|
||||
"无配料",
|
||||
"未见配料",
|
||||
)
|
||||
if any(x in t for x in refuse):
|
||||
return False
|
||||
# 真配料表通常含分隔符或足够长；避免「无」「暂无」等被当成命中
|
||||
if t in ("无", "暂无", "没有", "无。", "无，"):
|
||||
return False
|
||||
if _looks_like_recipe_or_dish_prep(t):
|
||||
return False
|
||||
if not _has_packaged_ingredient_table_signals(t):
|
||||
return False
|
||||
tail = _split_ingredient_segments(t)
|
||||
if len(tail) >= 32:
|
||||
if len(set(tail[-32:])) <= 3:
|
||||
return False
|
||||
sep_chars = "，、,;；"
|
||||
if len(t) < 18 and not any(c in t for c in sep_chars):
|
||||
return False
|
||||
return True
|
||||
|
||||
|
||||
REASON_NO_BODY_URLS = "【未识别到配料】未解析到任何详情长图 URL。"
|
||||
REASON_NO_VISION_API = (
|
||||
"【未识别到配料】未配置多模态 API（需环境变量 OPENAI_API_KEY + OPENAI_BASE_URL，"
|
||||
"或 LLM_API_KEY + LLM_BASE_URL）。"
|
||||
)
|
||||
|
||||
|
||||
def extract_ingredients_from_body_image_urls_reversed_with_source(
|
||||
urls_joined: str,
|
||||
*,
|
||||
referer: str | None = None,
|
||||
user_prompt: str | None = None,
|
||||
prompt_default: str | None = None,
|
||||
temperature: float | None = None,
|
||||
max_tokens: int | None = None,
|
||||
extra_json: dict[str, Any] | None = None,
|
||||
) -> tuple[str, str | None]:
|
||||
"""
|
||||
与 ``extract_ingredients_from_body_image_urls_reversed`` 相同逻辑；额外返回命中配料时所用的**图片 URL**
|
||||
（自后向前首次通过校验的那张）。未命中或失败时第二项为 ``None``。
|
||||
"""
|
||||
urls = parse_joined_image_urls(urls_joined)
|
||||
if not urls:
|
||||
return REASON_NO_BODY_URLS, None
|
||||
try:
|
||||
_resolve_credentials(None, None, None)
|
||||
except ValueError:
|
||||
return REASON_NO_VISION_API, None
|
||||
|
||||
ref = (referer if referer is not None else _d.IMAGE_REFERER) or "https://www.jd.com/"
|
||||
temp = float(temperature) if temperature is not None else float(_d.TEMPERATURE)
|
||||
mt = int(max_tokens) if max_tokens is not None else int(_d.MAX_TOKENS)
|
||||
extra = extra_json
|
||||
if extra is None and _d.QWEN_OMNI_TEMPLATE:
|
||||
extra = {"chat_template_kwargs": {"enable_thinking": False}}
|
||||
|
||||
pu = user_prompt if user_prompt is not None else ((_d.USER_PROMPT or "").strip() or None)
|
||||
pd = prompt_default if prompt_default is not None else _d.PROMPT_DEFAULT
|
||||
|
||||
n = len(urls)
|
||||
n_err = 0
|
||||
n_rejected = 0
|
||||
for url in reversed(urls):
|
||||
try:
|
||||
text = extract_ingredients_from_image(
|
||||
url,
|
||||
user_prompt=pu,
|
||||
referer=ref.strip(),
|
||||
temperature=temp,
|
||||
max_tokens=mt,
|
||||
extra_json=extra,
|
||||
prompt_default=pd,
|
||||
)
|
||||
except Exception:
|
||||
n_err += 1
|
||||
continue
|
||||
t = (text or "").strip()
|
||||
if _ingredient_extraction_acceptable(t):
|
||||
return t, url
|
||||
if t:
|
||||
n_rejected += 1
|
||||
|
||||
parts = [
|
||||
f"【未识别到配料】已对 {n} 张详情长图自后向前依次尝试（命中即停），未得到有效配料表。"
|
||||
]
|
||||
if n_err:
|
||||
parts.append(f" 请求异常 {n_err} 次。")
|
||||
if n_rejected:
|
||||
parts.append(f" 有 {n_rejected} 次返回未通过配料校验。")
|
||||
if not n_err and not n_rejected:
|
||||
parts.append(" 模型返回均为空或过短。")
|
||||
return "".join(parts), None
|
||||
|
||||
|
||||
def extract_ingredients_from_body_image_urls_reversed(
|
||||
urls_joined: str,
|
||||
*,
|
||||
referer: str | None = None,
|
||||
user_prompt: str | None = None,
|
||||
prompt_default: str | None = None,
|
||||
temperature: float | None = None,
|
||||
max_tokens: int | None = None,
|
||||
extra_json: dict[str, Any] | None = None,
|
||||
) -> str:
|
||||
"""
|
||||
对 URL 串拆出的链接 **从后往前**依次调用视觉模型：**首次**通过校验的配料文本立即返回（省时间）。
|
||||
|
||||
若始终无命中：返回以 ``【未识别到配料】`` 开头的原因说明（**不再返回空串**）。
|
||||
未配置 API 时返回 ``REASON_NO_VISION_API``。
|
||||
|
||||
命中条件（见 ``_ingredient_extraction_acceptable``）：须像**包装配料表**——「配料/含量」标题结构、
|
||||
``××（含量≥x%）``、或多段工业化原料逗号/顿号枚举（模型常省略标题）；纯家常备料（鸡胸、黄瓜、葱花等）
|
||||
仍丢弃并试下一张图。
|
||||
|
||||
若需同时得到所用图片 URL，请用 ``extract_ingredients_from_body_image_urls_reversed_with_source``。
|
||||
"""
|
||||
text, _ = extract_ingredients_from_body_image_urls_reversed_with_source(
|
||||
urls_joined,
|
||||
referer=referer,
|
||||
user_prompt=user_prompt,
|
||||
prompt_default=prompt_default,
|
||||
temperature=temperature,
|
||||
max_tokens=max_tokens,
|
||||
extra_json=extra_json,
|
||||
)
|
||||
return text
|
||||
100
backend/pipeline/openai_gateway/text_chat.py
Normal file
100
backend/pipeline/openai_gateway/text_chat.py
Normal file
@ -0,0 +1,100 @@
|
||||
"""
|
||||
OpenAI 兼容 `chat/completions` 纯文本（system + user），与多模态配料识别共用 `OPENAI_*` / `LLM_*` 环境配置。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import os
|
||||
from typing import Any
|
||||
|
||||
import requests
|
||||
|
||||
from .chat_content import normalize_message_content
|
||||
from .credentials import _resolve_credentials, resolve_text_model_name
|
||||
from .estimate import estimate_chat_input_tokens
|
||||
from .timeouts import chat_completion_read_timeout as _chat_completion_timeout
|
||||
|
||||
|
||||
def strip_outer_markdown_fence(text: str) -> str:
|
||||
"""若模型用 ``` / ```markdown 包裹全文，去掉最外层围栏。"""
|
||||
t = (text or "").strip()
|
||||
if not t.startswith("```"):
|
||||
return t
|
||||
lines = t.split("\n")
|
||||
if lines and lines[0].strip().startswith("```"):
|
||||
lines = lines[1:]
|
||||
while lines and lines[-1].strip() == "```":
|
||||
lines = lines[:-1]
|
||||
return "\n".join(lines).strip()
|
||||
|
||||
|
||||
def chat_completion_text(
|
||||
*,
|
||||
system_prompt: str,
|
||||
user_prompt: str,
|
||||
api_key: str | None = None,
|
||||
base_url: str | None = None,
|
||||
model: str | None = None,
|
||||
temperature: float = 0.2,
|
||||
max_tokens: int = 8192,
|
||||
timeout: int | tuple[float, float] | None = None,
|
||||
extra_json: dict[str, Any] | None = None,
|
||||
) -> str:
|
||||
if timeout is None:
|
||||
timeout = _chat_completion_timeout()
|
||||
k, b, _ = _resolve_credentials(api_key, base_url, None)
|
||||
m = resolve_text_model_name(model)
|
||||
body: dict[str, Any] = {
|
||||
"model": m,
|
||||
"messages": [
|
||||
{"role": "system", "content": system_prompt},
|
||||
{"role": "user", "content": user_prompt},
|
||||
],
|
||||
"temperature": temperature,
|
||||
"max_tokens": max_tokens,
|
||||
}
|
||||
if extra_json:
|
||||
body.update(extra_json)
|
||||
ctx_raw = (
|
||||
os.environ.get("LLM_CONTEXT_WINDOW")
|
||||
or os.environ.get("OPENAI_CONTEXT_WINDOW")
|
||||
or "32768"
|
||||
).strip()
|
||||
try:
|
||||
context_window = max(4096, int(ctx_raw))
|
||||
except ValueError:
|
||||
context_window = 32768
|
||||
buf = 256
|
||||
input_est = estimate_chat_input_tokens(system_prompt, user_prompt)
|
||||
if input_est >= context_window - buf - 256:
|
||||
raise ValueError(
|
||||
f"提示词过长（估算输入约 {input_est} tokens，上下文上限 {context_window}），"
|
||||
"请缩小报告/摘要输入或换更大上下文的模型；也可设置环境变量 LLM_CONTEXT_WINDOW。"
|
||||
)
|
||||
avail = context_window - input_est - buf
|
||||
want = int(body.get("max_tokens") or max_tokens)
|
||||
body["max_tokens"] = max(256, min(want, max(avail, 256)))
|
||||
r = requests.post(
|
||||
f"{b}/chat/completions",
|
||||
headers={
|
||||
"Authorization": f"Bearer {k}",
|
||||
"Content-Type": "application/json",
|
||||
},
|
||||
json=body,
|
||||
timeout=timeout,
|
||||
)
|
||||
try:
|
||||
r.raise_for_status()
|
||||
except requests.HTTPError as e:
|
||||
snippet = ""
|
||||
if e.response is not None:
|
||||
snippet = (e.response.text or "")[:1200].replace("\r\n", "\n").replace("\n", " ")
|
||||
if snippet:
|
||||
raise requests.HTTPError(
|
||||
f"{e!s} | body: {snippet}",
|
||||
response=e.response,
|
||||
request=e.request,
|
||||
) from e
|
||||
raise
|
||||
data = r.json()
|
||||
msg = (data.get("choices") or [{}])[0].get("message") or {}
|
||||
return normalize_message_content(msg.get("content"))
|
||||
22
backend/pipeline/openai_gateway/timeouts.py
Normal file
22
backend/pipeline/openai_gateway/timeouts.py
Normal file
@ -0,0 +1,22 @@
|
||||
"""(连接, 读) 超时时长：供 ``chat/completions`` 与视觉请求使用。"""
|
||||
from __future__ import annotations
|
||||
|
||||
import os
|
||||
|
||||
|
||||
def chat_completion_read_timeout() -> tuple[float, float]:
|
||||
read = 600
|
||||
raw = (os.environ.get("LLM_CHAT_TIMEOUT") or os.environ.get("OPENAI_TIMEOUT") or "").strip()
|
||||
if raw:
|
||||
try:
|
||||
read = max(60, int(raw))
|
||||
except ValueError:
|
||||
pass
|
||||
conn = 30.0
|
||||
raw_c = (os.environ.get("LLM_CHAT_CONNECT_TIMEOUT") or "").strip()
|
||||
if raw_c:
|
||||
try:
|
||||
conn = max(5.0, float(raw_c))
|
||||
except ValueError:
|
||||
pass
|
||||
return (conn, float(read))
|
||||
Loading…
x
Reference in New Issue
Block a user