physics_repeat_check/hnsw_retrieval.py

import json
import pickle
import requests
import numpy as np
from fuzzywuzzy import fuzz
from sentence_transformers import util
from pprint import pprint

import config
from formula_process import formula_recognize
from comprehensive_score import Comprehensive_Score
from physical_quantity_extract import physical_quantity_extract

class HNSW():
    def __init__(self, data_process, logger=None):
        # 配置初始数据
        self.mongo_coll = config.mongo_coll
        self.vector_dim = config.vector_dim
        self.hnsw_retrieve_url = config.hnsw_retrieve_url
        self.dim_classify_url = config.dim_classify_url
        self.knowledge_tagging_url = config.knowledge_tagging_url
        # 日志采集
        self.logger = logger
        self.log_msg = config.log_msg
        # 数据预处理实例化
        self.dpp = data_process
        # 语义相似度实例化
        self.cph_score = Comprehensive_Score(config.dev_mode)
        # 难度数值化定义
        self.difficulty_transfer = {"容易": 0.2, "较易": 0.4, "一般": 0.6, "较难": 0.8, "困难": 1.0}
        # 加载公式处理数据模型(词袋模型/原始向量/原始数据)
        with open(config.bow_model_path, "rb") as bm:
            self.bow_model = pickle.load(bm)
        self.bow_vector = np.load(config.bow_vector_path)
        with open(config.formula_data_path, 'r', encoding='utf8', errors='ignore') as f:
            self.formula_id_list = json.load(f)

    # 图片搜索查重功能
    def img_retrieve(self, retrieve_text, post_url, similar, topic_num):
        try:
            if post_url is not None:
                # 日志采集
                if self.logger is not None:
                    self.logger.info(self.log_msg.format(id="图片搜索查重",
                                                          type="{}图片搜索查重post".format(topic_num),
                                                          message=retrieve_text))
                img_dict = dict(img_url=retrieve_text, img_threshold=similar, img_max_num=40)
                img_res = requests.post(post_url, json=img_dict, timeout=30).json()
                # 日志采集
                if self.logger is not None:
                    self.logger.info(self.log_msg.format(id="图片搜索查重",
                                                          type="{}图片搜索查重success".format(topic_num),
                                                          message=img_res))
                return img_res
        except Exception as e:
            # 日志采集
            if self.logger is not None:
                self.logger.error(self.log_msg.format(id="图片搜索查重",
                                                        type="{}图片搜索查重error".format(topic_num),
                                                        message=retrieve_text))
            return []
    
    # 公式搜索查重功能
    def formula_retrieve(self, formula_string, similar):
        # 调用清洗分词函数
        formula_string = '$' + formula_string + '$'
        formula_string = self.dpp.content_clear_func(formula_string)
        # 公式识别
        formula_list = formula_recognize(formula_string)
        if len(formula_list) == 0:
            return []
        # 日志采集
        if self.logger is not None:
            self.logger.info(self.log_msg.format(id="formula_retrieve",
                                                 type=formula_string,
                                                 message=formula_list))
        try:
            # 使用词袋模型计算句向量
            bow_vec = self.bow_model.transform([formula_list[0]]).toarray().astype("float32")
            # 并行计算余弦相似度
            formula_cos = np.array(util.cos_sim(bow_vec, self.bow_vector))
            # 获取余弦值大于等于0.8的数据索引
            cos_list = np.where(formula_cos[0] >= similar)[0]
        except:
            return []
        if len(cos_list) == 0:
            return []
        # 根据阈值获取满足条件的题目id
        res_list = []
        # formula_threshold = similar
        formula_threshold = 0.7
        for idx in cos_list:
            fuzz_score = fuzz.ratio(formula_list[0], self.formula_id_list[idx][0]) / 100
            if fuzz_score >= formula_threshold:
                # res_list.append([self.formula_id_list[idx][1], fuzz_score])
                # 对余弦相似度进行折算
                cosine_score = formula_cos[0][idx]
                if 0.95 <= cosine_score < 0.98:
                    cosine_score = cosine_score * 0.98
                elif cosine_score < 0.95:
                    cosine_score = cosine_score * 0.95
                # 余弦相似度折算后阈值判断
                if cosine_score < similar:
                    continue
                res_list.append([self.formula_id_list[idx][1], int(cosine_score * 100) / 100])
        # 根据分数对题目id排序并返回前50个
        res_sort_list = sorted(res_list, key=lambda x: x[1], reverse=True)[:80]
        formula_res_list = []
        fid_set = set()
        for ele in res_sort_list:
            for fid in ele[0]:
                if fid in fid_set:
                    continue
                fid_set.add(fid)
                formula_res_list.append([fid, ele[1]])

        return formula_res_list[:50]

    def api_post(self, post_url, post_data, log_info):
        try:
            post_result = requests.post(post_url, json=post_data, timeout=10).json()
        except Exception as e:
            post_result = []
            # 日志采集
            if self.logger is not None:
                self.logger.error(self.log_msg.format(id="{}error".format(log_info[0]),
                                                        type="当前题目{}error".format(log_info[0]),
                                                        message=log_info[1]))
        
        return post_result

    # HNSW查(支持多学科混合查重)
    def retrieve(self, retrieve_list, post_url, similar, scale, doc_flag):
        """
        return: 
        [
            {
                'semantics': [[20232015, 1.0, {'quesType': 1.0, 'knowledge': 1.0, 'physical_scene': 1.0, 'solving_type': 1.0, 'difficulty': 1.0, 'physical_quantity': 1.0}]], 
                'text': [[20232015, 0.97]], 
                'image': [], 
                'label': {'knowledge': ['串并联电路的辨别'], 'physical_scene': ['串并联电路的辨别'], 'solving_type': ['规律理解'], 'difficulty': 0.6, 'physical_quantity': ['电流']}, 
                'topic_num': 1
            }, 
            ...
        ]
        """
        # 计算retrieve_list的vec值
        # 调用清洗分词函数和句向量计算函数
        sent_vec_list, cont_clear_list = self.dpp(retrieve_list, is_retrieve=True)
        # HNSW查重
        retrieve_res_list = []
        for i,sent_vec in enumerate(sent_vec_list):
            # 初始化返回数据类型
            # retrieve_value_dict = dict(synthese=[], semantics=[], text=[], image=[])
            retrieve_value_dict = dict(semantics=[], text=[], image=[])
            # 获取题目序号
            topic_num = retrieve_list[i]["topic_num"] if "topic_num" in retrieve_list[i] else 1
            # 图片搜索查重功能
            if doc_flag is True:
                retrieve_value_dict["image"] = self.img_retrieve(retrieve_list[i]["stem"], post_url, similar, topic_num)
            else:
                retrieve_value_dict["image"] = []
            """
            文本相似度特殊处理
            """
            # 判断句向量维度
            if sent_vec.size != self.vector_dim:
                retrieve_res_list.append(retrieve_value_dict)
                continue
            # 调用hnsw接口检索数据
            query_labels = self.api_post(self.hnsw_retrieve_url, sent_vec.tolist(), ["HNSW检索", cont_clear_list[i]])
            if len(query_labels) == 0:
                retrieve_res_list.append(retrieve_value_dict)
                continue
            # 批量读取数据库
            mongo_find_dict = {"id": {"$in": query_labels}}
            query_dataset = self.mongo_coll.find(mongo_find_dict)
            ####################################### 语义相似度借靠 ####################################### 
            query_data = dict()
            ####################################### 语义相似度借靠 ####################################### 
            # 返回大于阈值的结果
            for label_data in query_dataset:
                if "sentence_vec" not in label_data:
                    continue
                # 计算余弦相似度得分
                label_vec = pickle.loads(label_data["sentence_vec"])
                if label_vec.size != self.vector_dim:
                    continue
                cosine_score = util.cos_sim(sent_vec, label_vec)[0][0]
                # 阈值判断
                if cosine_score < similar:
                    continue
                # 计算编辑距离得分
                fuzz_score = fuzz.ratio(cont_clear_list[i], label_data["content_clear"]) / 100
                # 进行编辑距离得分验证,若小于设定分则过滤
                if fuzz_score >= similar:
                    retrieve_value = [label_data["id"], fuzz_score]
                    retrieve_value_dict["text"].append(retrieve_value)
                    ####################################### 语义相似度借靠 ####################################### 
                    max_mark_score = 0
                    if cosine_score >= 0.9 and cosine_score > max_mark_score:
                        max_mark_score = cosine_score
                        query_data = label_data
                    ####################################### 语义相似度借靠 ####################################### 
            
            """
            语义相似度特殊处理
            """
            # 标签字典初始化
            label_dict = dict()  # todo: 可以加一个高相似的替换, 不走模型预测
            # 知识点LLM标注
            label_dict["knowledge"] = query_data["knowledge"] if query_data else []
            ####################################### 知识点标注接口调用 ####################################### 
            # knowledge_post_data = {"sentence": cont_clear_list[i]}
            # label_dict["knowledge"] = self.api_post(self.knowledge_tagging_url, knowledge_post_data, ["知识点标注", cont_clear_list[i]])
            ####################################### 知识点标注接口调用 ####################################### 
            tagging_id_list = [self.cph_score.knowledge2id[ele] for ele in label_dict["knowledge"] \
                               if ele in self.cph_score.knowledge2id]
            # 题型数据获取
            label_dict["quesType"] = retrieve_list[i].get("quesType", "选择题")
            # 多维分类api调用
            dim_post_data = {"sentence": cont_clear_list[i], "quesType": label_dict["quesType"]}
            dim_classify_dict = self.api_post(self.dim_classify_url, dim_post_data, ["多维分类", cont_clear_list[i]])
            if len(dim_classify_dict) == 0:
                dim_classify_dict = {"solving_type": ["规律理解"], "difficulty": 0.6}
            # 求解类型模型分类
            label_dict["solving_type"] = dim_classify_dict["solving_type"]
            # 难度模型分类
            label_dict["difficulty"] = dim_classify_dict["difficulty"]
            # 物理量规则提取
            label_dict["physical_quantity"] = physical_quantity_extract(cont_clear_list[i])

            # LLM标注知识点题目获取题库对应相似知识点题目数据
            knowledge_id_list = []
            if len(tagging_id_list) > 0:
                ####################################### encode_base_value设置 ####################################### 
                # 考试院: 10000, 风向标: 10
                encode_base_value = 10000 if config.dev_mode == "ksy" else 10
                ####################################### encode_base_value设置 ####################################### 
                for ele in tagging_id_list:
                    init_id = int(ele / encode_base_value) * encode_base_value
                    init_id_list = self.cph_score.init_id2max_id.get(str(init_id), [])
                    knowledge_id_list.extend(init_id_list)
            knowledge_query_dataset = None
            if len(knowledge_id_list) > 0:
                mongo_find_dict = {"knowledge_id": {"$in": knowledge_id_list}}
                knowledge_query_dataset = self.mongo_coll.find(mongo_find_dict)
            # 返回大于阈值的结果
            if knowledge_query_dataset:
                # 待查重难度数值转换
                if label_dict["difficulty"] in self.difficulty_transfer:
                    label_dict["difficulty"] = self.difficulty_transfer[label_dict["difficulty"]]
                for refer_data in knowledge_query_dataset:
                    # 题库数据难度数值转换
                    if refer_data["difficulty"] in self.difficulty_transfer:
                        refer_data["difficulty"] = self.difficulty_transfer[refer_data["difficulty"]]
                    sum_score, score_dict = self.cph_score(label_dict, refer_data, scale)
                    if sum_score < similar:
                        continue
                    retrieve_value = [refer_data["id"], sum_score, score_dict]
                    retrieve_value_dict["semantics"].append(retrieve_value)

            # 将组合结果按照score降序排序
            retrieve_sort_dict = {k: sorted(value, key=lambda x: x[1], reverse=True)
                                  for k,value in retrieve_value_dict.items()}

            # 加入题目序号
            retrieve_sort_dict["label"] = label_dict
            retrieve_sort_dict["topic_num"] = topic_num
            
            # 以字典形式返回最终查重结果
            retrieve_res_list.append(retrieve_sort_dict)

        return retrieve_res_list


if __name__ == "__main__":
    # 获取mongodb数据
    mongo_coll = config.mongo_coll
    from data_preprocessing import DataPreProcessing
    hnsw = HNSW(DataPreProcessing())

    test_data = []
    for idx in [201511100736265]:
        test_data.append(mongo_coll.find_one({"id": idx}))

    res = hnsw.retrieve(test_data, '', 0.8, False)
    pprint(res[0]["semantics"])

    # # 公式搜索查重功能
    # formula_string = "ρ蜡=0.9*10^3Kg/m^3"
    # formula_string = "p蜡=0.9*10^3Kq/m^3"
    # print(hnsw.formula_retrieve(formula_string, 0.8))