ElasticSearch rank_eval报错：Failed to build [request] after last required field arrived

在LinkSO数据集上使用ElasticSearch开展排序评估工作时，调用es.rank_eval()函数触发400错误：

BadRequestError: BadRequestError(400, 'x_content_parse_exception', 'Failed to build [request] after last required field arrived')

相关代码及配置如下：

1. 获取NDCG的get_ndcg函数 #

def get_ndcg(dataframe, input_index="java_lm"):
    ndcg_list = []
    # loop through each qid1
    for i in range(0, len(dataframe["qid1"]), 30):
        qid1_title = es.get(index=input_index, id=dataframe["qid1"][i])['_source']['title']
        
        # load ratings from the json file
        f = open("qids/" + input_index + "/" + str(dataframe["qid1"][i]) + ".json")
        data = json.load(f)

        _search = ranking(dataframe["qid1"][i], qid1_title, ratings=data)
        
        result = es.rank_eval(index=input_index, body=_search)
        
        ndcg = result['metric_score']
        ndcg_list.append(ndcg)
        
    return ndcg_list

2. 生成排序请求的ranking函数 #

def ranking(qid1, qid1_title, ratings):
    _search = {
        "requests": [
            {
            "id": str(qid1),
            "request": {
                "query": {
                    "bool": {
                        "must_not": {
                            "match": {
                                "_id": qid1
                            }
                        },
                        "should": [
                            {
                                "match": {
                                    "title": {
                                        "query": qid1_title,
                                        "boost": 3.0,
                                        "analyzer": "my_analyzer"
                                    }
                                }
                            },
                            {
                                "match": {
                                    "body": {
                                        "query": qid1_title,
                                        "boost": 0.5,
                                        "analyzer": "my_analyzer"
                                    }
                                }
                            },
                            {
                                "match": {
                                    "answer": {
                                        "query": qid1_title,
                                        "boost": 0.5,
                                        "analyzer": "my_analyzer"
                                    }
                                }
                            }
                        ]
                    }
                }
            },
            "ratings": ratings
            }
        ],
        "metric": {
            "dcg": {
                "k": 10,
                "normalize": True
            }
        }
    }
    return _search

3. ratings文件格式 #

[
{"_index": "java_lm", "_id": "15194804", "rating": 0},
{"_index": "java_lm", "_id": "18264178", "rating": 0},
{"_index": "java_lm", "_id": "16225177", "rating": 1},
{"_index": "java_lm", "_id": "16445238", "rating": 0},
{"_index": "java_lm", "_id": "17233226", "rating": 0}
]

4. Kibana创建索引的PUT命令 #

PUT /java_lm
{
  "settings": {
    "similarity": {
      "LM": {
        "type": "LMDirichlet",
        "mu": 2000
      }
    },
    "analysis": {
      "analyzer": {
        "my_analyzer": {
          "tokenizer": "whitespace",
          "filter": [
            "lowercase",
            "porter_stem"
          ]
        }
      }
    }
  },
  "mappings": {
    "properties": {
      "title": {
        "type": "text",
        "analyzer": "my_analyzer",
        "similarity": "LM"
      },
      "body": {
        "type": "text",
        "analyzer": "my_analyzer",
        "similarity": "LM"
      },
      "answer": {
        "type": "text",
        "analyzer": "my_analyzer",
        "similarity": "LM"
      }
    }
  }
}

核心错误原因 #

1. Metric字段格式不兼容：使用dcg类型并指定normalize: true的写法在部分Elasticsearch版本中存在解析问题，官方更推荐直接使用ndcg类型（默认自动归一化）。
2. _id查询方式错误：_id是关键字字段，match查询会对输入分词，导致无法精准匹配，应该用term查询。
3. 文件读取未规范关闭：直接open文件后未关闭，可能导致资源泄漏或读取异常。

具体修正步骤 #

1. 修正Metric配置 #

将metric部分改为ndcg类型，无需额外指定归一化参数：

"metric": {
    "ndcg": {
        "k": 10
    }
}

2. 修正must_not子句的查询方式 #

用term查询替代match查询_id，确保精准匹配：

"must_not": {
    "term": {
        "_id": str(qid1)  # 强制转为字符串，避免类型不匹配
    }
}

3. 规范文件读取方式 #

使用上下文管理器with自动关闭文件，避免资源问题：

with open(f"qids/{input_index}/{str(dataframe['qid1'][i])}.json", "r") as f:
    data = json.load(f)

4. 完整修正后的ranking函数 #

def ranking(qid1, qid1_title, ratings):
    _search = {
        "requests": [
            {
                "id": str(qid1),
                "request": {
                    "query": {
                        "bool": {
                            "must_not": {
                                "term": {
                                    "_id": str(qid1)
                                }
                            },
                            "should": [
                                {
                                    "match": {
                                        "title": {
                                            "query": qid1_title,
                                            "boost": 3.0,
                                            "analyzer": "my_analyzer"
                                        }
                                    }
                                },
                                {
                                    "match": {
                                        "body": {
                                            "query": qid1_title,
                                            "boost": 0.5,
                                            "analyzer": "my_analyzer"
                                        }
                                    }
                                },
                                {
                                    "match": {
                                        "answer": {
                                            "query": qid1_title,
                                            "boost": 0.5,
                                            "analyzer": "my_analyzer"
                                        }
                                    }
                                }
                            ]
                        }
                    }
                },
                "ratings": ratings
            }
        ],
        "metric": {
            "ndcg": {
                "k": 10
            }
        }
    }
    return _search

5. 额外优化：循环逻辑合理性 #

原循环步长30可能跳过部分qid1，建议改为遍历所有唯一qid1：

for qid1 in dataframe["qid1"].unique():
    qid1_title = es.get(index=input_index, id=qid1)['_source']['title']
    # 后续读取文件、生成请求逻辑不变

内容的提问来源于stack exchange，提问作者insanely_a_