feat: 新增物料/入库单实时 CLIP 向量提取（新建+更新），修复 I/O 延迟和路径解析静默失败

inventory-backend/app/models/base.py

@@ -1,5 +1,6 @@
 # app/models/base.py
 from app.extensions import db
+from pgvector.sqlalchemy import Vector
 import json
 
 
@@ -34,6 +35,9 @@ class MaterialBase(db.Model):
     # 强制质检标记（采购入库时必须上传检测报告）
     is_inspection_required = db.Column(db.Boolean, default=False, comment='是否强制要求质检')
 
+    # CLIP 视觉向量（用于以图搜图）
+    img_embedding = db.Column(Vector(512), nullable=True)
+
     # ============================================================
     # 关联关系区域
     # ============================================================

inventory-backend/app/models/inbound/buy.py

@@ -1,5 +1,6 @@
 # inventory-backend/app/models/inbound/buy.py
 from app.extensions import db
+from pgvector.sqlalchemy import Vector
 import json
 # 显式导入 MaterialBase 以防 relationship 找不到引用
 from app.models.base import MaterialBase
@@ -55,6 +56,9 @@ class StockBuy(db.Model):
     # 全局打印流水号
     global_print_id = db.Column(db.Integer)
 
+    # CLIP 视觉向量（用于以图搜图）
+    arrival_image_embedding = db.Column(Vector(512), nullable=True)
+
     # 关系定义
     base = db.relationship('MaterialBase', back_populates='stock_buys')
 

inventory-backend/app/models/inbound/product.py

@@ -1,5 +1,6 @@
 # app/models/inbound/product.py
 from app.extensions import db
+from pgvector.sqlalchemy import Vector
 import json
 from app.models.base import MaterialBase
 
@@ -58,6 +59,9 @@ class StockProduct(db.Model):
     # 全局打印流水号
     global_print_id = db.Column(db.Integer)
 
+    # CLIP 视觉向量（用于以图搜图）
+    arrival_image_embedding = db.Column(Vector(512), nullable=True)
+
     # 关系定义
     base = db.relationship('MaterialBase', back_populates='stock_products')
 

inventory-backend/app/models/inbound/semi.py

@@ -1,5 +1,6 @@
 # app/models/inbound/semi.py
 from app.extensions import db
+from pgvector.sqlalchemy import Vector
 import json
 from app.models.base import MaterialBase
 
@@ -56,6 +57,9 @@ class StockSemi(db.Model):
     # 全局打印流水号
     global_print_id = db.Column(db.Integer)
 
+    # CLIP 视觉向量（用于以图搜图）
+    arrival_image_embedding = db.Column(Vector(512), nullable=True)
+
     # 关系定义
     base = db.relationship('MaterialBase', back_populates='stock_semis')
 

inventory-backend/app/services/inbound/base_service.py

@@ -12,6 +12,7 @@ import traceback
 import json
 import io
 import datetime
+from app.utils.ai_vision import extract_and_embed
 # 需要 pip install openpyxl
 from openpyxl import Workbook
 from openpyxl.styles import Font, Alignment, Border, Side, PatternFill
@@ -555,7 +556,9 @@ class MaterialBaseService:
                 product_image=json.dumps(data.get('generalImage', [])),
                 is_enabled=is_enabled_val
             )
-
+            # 实时提取产品图向量（失败不影响业务）
+            if new_material.product_image:
+                new_material.img_embedding = extract_and_embed(new_material.product_image)
             db.session.add(new_material)
             db.session.commit()
             return new_material
@@ -587,6 +590,10 @@ class MaterialBaseService:
             if 'generalImage' in data:
                 material.product_image = json.dumps(data['generalImage'])
 
+                # 补上这两行：提取新上传图片的向量！
+                if material.product_image:
+                    material.img_embedding = extract_and_embed(material.product_image)
+
             # 【核心修改】：兼容前端传来的布尔值
             if 'isEnabled' in data:
                 raw_enabled = data['isEnabled']

inventory-backend/app/services/inbound/buy_service.py

@@ -9,6 +9,7 @@ from sqlalchemy import or_, func, text, and_
 from sqlalchemy.exc import IntegrityError
 import traceback
 import json
+from app.utils.ai_vision import extract_and_embed
 
 
 class BuyInboundService:
@@ -177,6 +178,9 @@ class BuyInboundService:
                 arrival_photo=json.dumps(data.get('arrival_photo', [])),
                 inspection_report=json.dumps(data.get('inspection_report', []))
             )
+            # 实时提取到货图片向量（失败不影响业务）
+            if new_stock.arrival_photo:
+                new_stock.arrival_image_embedding = extract_and_embed(new_stock.arrival_photo)
             db.session.add(new_stock)
             db.session.commit()
             return new_stock

inventory-backend/app/services/inbound/product_service.py

@@ -9,6 +9,7 @@ from sqlalchemy import or_, func, text, and_
 from sqlalchemy.exc import IntegrityError
 import traceback
 import json
+from app.utils.ai_vision import extract_and_embed
 
 
 class ProductInboundService:
@@ -183,6 +184,9 @@ class ProductInboundService:
                 sale_price=float(data.get('sale_price') or 0),
                 order_id=data.get('order_id')
             )
+            # 实时提取成品实拍图向量（失败不影响业务）
+            if new_stock.product_photo:
+                new_stock.arrival_image_embedding = extract_and_embed(new_stock.product_photo)
             db.session.add(new_stock)
             db.session.commit()
             return new_stock

inventory-backend/app/services/inbound/semi_service.py

@@ -9,6 +9,7 @@ from sqlalchemy import or_, func, text, and_
 from sqlalchemy.exc import IntegrityError
 import traceback
 import json
+from app.utils.ai_vision import extract_and_embed
 
 
 class SemiInboundService:
@@ -220,6 +221,9 @@ class SemiInboundService:
                 detail_link=data.get('detail_link'),
                 remark=data.get('remark')
             )
+            # 实时提取到货图片向量（失败不影响业务）
+            if new_stock.arrival_photo:
+                new_stock.arrival_image_embedding = extract_and_embed(new_stock.arrival_photo)
             db.session.add(new_stock)
             db.session.commit()
             return new_stock

inventory-backend/app/utils/ai_vision.py

@@ -4,6 +4,8 @@ AI Vision 模块 - CLIP Vision Encoder ONNX 推理
 """
 
 import os
+import json
+import time
 import numpy as np
 from PIL import Image
 import onnxruntime as ort
@@ -130,3 +132,61 @@ def get_image_embedding(image_path: str) -> list:
         }
     )
     return outputs[0][0].tolist()
+
+
+# ============================================================================
+# 通用向量提取工具：防呆、防错
+# ============================================================================
+
+def extract_and_embed(photo_source):
+    if not photo_source:
+        return None
+    try:
+        # 1. 提取基础字符串
+        photo_source_str = str(photo_source).strip()
+        raw_path = ""
+
+        # 尝试剥掉 JSON 外壳
+        try:
+            parsed = json.loads(photo_source_str)
+            if isinstance(parsed, list):
+                raw_path = parsed[0] if parsed else ""
+            elif isinstance(parsed, str):
+                raw_path = parsed
+            else:
+                raw_path = str(parsed)
+        except:
+            raw_path = photo_source_str
+
+        if not raw_path:
+            return None
+
+        # 2. 剥离出最纯净的文件名 (只取最后一段)
+        pure_filename = raw_path.split('/')[-1]
+
+        # 3. 【终极物理净化】强行抠掉所有多余的标点符号！
+        # 哪怕传进来的是 123.jpg"] 或者是 "123.jpg"，全部洗干净
+        pure_filename = pure_filename.replace('"', '').replace("'", "").replace('[', '').replace(']', '')
+
+        # 4. 拼接真实的 Docker 物理路径
+        file_path = os.path.join('/app/uploads', pure_filename)
+
+        # 5. 加入重试机制 (最多等 3 秒)
+        max_retries = 6
+        for i in range(max_retries):
+            if os.path.exists(file_path):
+                # 文件找到了，开始提取向量
+                vec = get_image_embedding(file_path)
+                if isinstance(vec, np.ndarray):
+                    return vec.tolist()
+                return vec
+            else:
+                print(f"[AI 识图等待] 第 {i+1} 次尝试，未找到文件 {file_path}，等待 0.5s...")
+                time.sleep(0.5)
+
+        print(f"[AI 识图警告] 彻底失败！经过等待依然未找到图片: {file_path}")
+
+    except Exception as e:
+        print(f"[AI 识图错误] 实时提取向量失败: {str(e)}")
+
+    return None

inventory-backend/requirements.txt

@@ -28,3 +28,5 @@ APScheduler==3.10.4
 pytz
 # [新增] 进度条库 (脚本和任务所需)
 tqdm>=4.66.0
+# [新增] pgvector 向量数据库支持（以图搜图 / 实时向量提取）
+pgvector>=0.2.0