KCGL/inventory-backend/app/api/v1/common/image_search.py

# -*- coding: utf-8 -*-
"""
以图搜图 API - CLIP Vision Embedding + pgvector 余弦距离检索
"""

import os
import uuid
import json
from flask import Blueprint, request, jsonify
from sqlalchemy import text
from app.extensions import db
from app.utils.ai_vision import load_clip_model, get_image_embedding

# 注册蓝图
image_search_bp = Blueprint('image_search', __name__)


# ============================================================================
# POST /api/v1/common/image-search
# 以图搜图：上传图片 → CLIP embedding → pgvector 余弦相似度检索
# ============================================================================

@image_search_bp.route('/image-search', methods=['POST'])
def image_search():
    # ---------------------------------------------------------
    # 1. 检查文件
    # ---------------------------------------------------------
    if 'file' not in request.files:
        return jsonify({"code": 400, "msg": "未找到图片文件"}), 400

    file = request.files['file']
    if file.filename == '':
        return jsonify({"code": 400, "msg": "未选择文件"}), 400

    # ---------------------------------------------------------
    # 2. 安全保存临时文件
    # ---------------------------------------------------------
    ext = file.filename.rsplit('.', 1)[-1].lower()
    if ext not in {'png', 'jpg', 'jpeg', 'gif', 'bmp', 'webp'}:
        return jsonify({"code": 400, "msg": "不支持的图片格式"}), 400

    tmp_filename = f"{uuid.uuid4().hex}.{ext}"
    tmp_dir = os.path.join(os.path.dirname(__file__), '..', '..', '..', 'uploads')
    os.makedirs(tmp_dir, exist_ok=True)
    tmp_path = os.path.join(tmp_dir, tmp_filename)

    try:
        file.save(tmp_path)
        print(f"💾 [ImageSearch] 临时文件已保存: {tmp_path}")

        # ---------------------------------------------------------
        # 3. 提取 CLIP embedding
        # ---------------------------------------------------------
        load_clip_model()
        embedding = get_image_embedding(tmp_path)
        print(f"✅ [ImageSearch] Embedding 提取成功，维度: {len(embedding)}")

    except Exception as e:
        print(f"❌ [ImageSearch] 图像处理失败: {e}")
        return jsonify({"code": 500, "msg": f"图像处理失败: {str(e)}"}), 500

    finally:
        # ---------------------------------------------------------
        # 4. 无论成功与否，都删除临时文件
        # ---------------------------------------------------------
        if os.path.exists(tmp_path):
            try:
                os.remove(tmp_path)
                print(f"🗑️ [ImageSearch] 临时文件已清理: {tmp_path}")
            except Exception as e:
                print(f"⚠️ [ImageSearch] 临时文件删除失败: {e}")

    # ---------------------------------------------------------
    # 5. pgvector 余弦相似度检索（跨表联合检索）
    # ---------------------------------------------------------
    try:
        query_vector_str = '[' + ','.join(str(v) for v in embedding) + ']'

        sql = text("""
    SELECT id, name, spec_model, image_url,
           (1 - (vec <=> :query_vector)) AS similarity
    FROM (
        -- 1. 基础物料表
        SELECT id, name, spec_model, product_image AS image_url, img_embedding AS vec
        FROM material_base 
        WHERE img_embedding IS NOT NULL
        
        UNION ALL
        
        -- 2. 采购入库表 (通过 base_id 关联拿真实物料信息)
        SELECT mb.id, mb.name, mb.spec_model, sb.arrival_photo AS image_url, sb.arrival_image_embedding AS vec
        FROM stock_buy sb
        JOIN material_base mb ON sb.base_id = mb.id
        WHERE sb.arrival_image_embedding IS NOT NULL
        
        UNION ALL
        
        -- 3. 半成品入库表 (通过 base_id 关联拿真实物料信息)
        SELECT mb.id, mb.name, mb.spec_model, ss.arrival_photo AS image_url, ss.arrival_image_embedding AS vec
        FROM stock_semi ss
        JOIN material_base mb ON ss.base_id = mb.id
        WHERE ss.arrival_image_embedding IS NOT NULL
        
        UNION ALL
        
        -- 4. 成品入库表 (通过 base_id 关联拿真实物料信息)
        SELECT mb.id, mb.name, mb.spec_model, sp.product_photo AS image_url, sp.arrival_image_embedding AS vec
        FROM stock_product sp
        JOIN material_base mb ON sp.base_id = mb.id
        WHERE sp.arrival_image_embedding IS NOT NULL
    ) AS combined
    
    -- 核心：计算余弦距离并排序，取最接近的前 50 个！
    ORDER BY vec <=> :query_vector LIMIT 50
""")

        # 执行查询
        records = db.session.execute(sql, {"query_vector": query_vector_str}).fetchall()

        results = []
        seen_product_ids = set()  # 【新增】用来记录已经添加过的物料 ID

        for row in records:
            # 【新增】如果这个物料已经在这个列表里了，直接跳过它
            if row.id in seen_product_ids:
                continue

            # 记录这个物料 ID，保证下次不会再重复添加
            seen_product_ids.add(row.id)

            # 1. 提取原始 URL
            raw_url = row.image_url
            clean_url = ""

            if raw_url:
                if raw_url.startswith('[') and raw_url.endswith(']'):
                    import json
                    try:
                        url_list = json.loads(raw_url)
                        clean_url = url_list[0] if url_list else ""
                    except:
                        clean_url = raw_url
                else:
                    clean_url = raw_url

            # 2. 组装返回结果
            results.append({
                "product_id": row.id,
                "product_name": row.name,
                "spec_model": row.spec_model,
                "image_url": clean_url,
                "similarity": round(float(row.similarity), 4)
            })

            # 修改后：只要凑够了 10 个完全不同的物料，就立刻结束循环
            if len(results) >= 10:
                break

        return jsonify({"code": 200, "data": results})

    except Exception as e:
        print(f"❌ [ImageSearch] 数据库检索失败: {e}")
        return jsonify({"code": 500, "msg": f"检索失败: {str(e)}"}), 500