mirror of
https://github.com/primedigitaltech/market-assistant.git
synced 2026-07-21 23:41:39 +08:00
9fd32a3f84
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
536 lines
18 KiB
Python
536 lines
18 KiB
Python
"""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 |