我们测试了JoyAgent单表问数表现，整体问数效果高于预期，能够自动加工复杂的数据场景。例如：我们导入员工信息表，包含身份证号码和出生日期等字段，我们通过问数"获取身份证号码中的出生日期和登记的出生日期有差异的员工"，系统先从身份证号码中提取出生日期，并做日期格式转换，再判断是否相同，可以很好的反馈正确数据。
但是，真实分析场景很少在单张表上做数据分析，会横跨多张表做数据关联。JoyAgent的源代码以及样例中并未展示或测试，需要多表关联的业务问数。我们在检查源代码后，评估JoyAgent可以完成多表关联任务，但是，源代码并无外键关系知识库编写规范和样例可供参考。于是，我们设计了一个多表关联的数据场景。并在此场景下，做了工程测试。
我们在完成场景设计和测试数据导入后，系统自动生成了表关联SQL代码，但是在表名字替换阶段出现了问题。下面我们逐步描述我们设置的查询场景，以及查询场景之下遇到的一个系统bug，并成功修复，达到预期效果。我们逐步描述场景、问题、修复方案
1.数据场景
我们导入两张表作为问数基础：员工信息表，员工考勤信息表。建表语句如下：
CREATE TABLE employee_info (
employee_id VARCHAR(20) PRIMARY KEY COMMENT '员工ID（主键）',
full_name VARCHAR(50) NOT NULL COMMENT '员工全名',
gender VARCHAR(50) COMMENT '性别:男或女',
nationality VARCHAR(30) COMMENT '国籍',
id_card VARCHAR(20) COMMENT '身份证号',
birth_date VARCHAR(200) NOT NULL COMMENT '出生日期',
department VARCHAR(50) NOT NULL COMMENT '所属部门',
marital_status VARCHAR(20) COMMENT '婚姻状况（未婚,已婚,离异）',
education VARCHAR(20) COMMENT '最高学历（高中,专科,本科,硕士,博士）',
contact_phone VARCHAR(15) COMMENT '联系电话',
emergency_contact VARCHAR(15) COMMENT '紧急联系人电话',
address VARCHAR(1000) COMMENT '现居住地址',
hire_date VARCHAR(100) NOT NULL COMMENT '入职日期'
) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COMMENT='员工基础信息表';
CREATE TABLE employee_attendance (
attendance_id INT PRIMARY KEY COMMENT '行 ID',
employee_id VARCHAR(20) COMMENT '员工ID',
clock_in_time VARCHAR(20) COMMENT '上班打卡时间',
clock_out_time VARCHAR(20) COMMENT '下班打卡时间',
clock_in_type varchar(100) COMMENT '打卡方式（指纹,人脸识别, IC卡, 手机APP, 手动补录)'
) COMMENT='员工打卡信息表';数据模型配置：
- name: 员工基础信息表
id: t_employeeinfo
type: table
content: employee_info
remark: 员工基础信息表
business-prompt:
ignore-fields:
default-recall-fields:
analyze-suggest-fields:
analyze-forbid-fields: employee_id
sync-value-fields:
column-alias-map: ''
- name: 员工考勤明细表
id: t_employeeattendance
type: table
content: employee_attendance
remark: 包含员工的考勤打卡记录
business-prompt:
ignore-fields:
default-recall-fields:
analyze-suggest-fields:
analyze-forbid-fields: attendance_id
sync-value-fields:
column-alias-map: ''测试数据（略），可以通过deepseek按照这种数据模式帮助造几条数据。
2.系统异常分析
当我们问出"获取考勤打卡迟到员工的姓名"，可以正常生成SQL伪代码，SQL中的表名称是配置的知识库ID：
2025-10-17T03:25:44.730Z INFO 47 --- genie-backend 48.237:1601/... com.jd.genie.service.Nl2SqlService : b44bf49f-e299-4c17-8e72-90b991325a6e,b44bf49f-e299-4c17-8e72-90b991325a6e SSE数据结果：{"code": 200, "data": {"query": "通过员工信息表和考勤记录表，获取每个员工的迟到情况", "nl2sql": "**SELECT `t_employeeinfo`.`EMPLOYEE_ID`, `t_employeeinfo`.`FULL_NAME`, COUNT(CASE WHEN CAST(`t_employeeattendance`.`CLOCK_IN_TIME` AS TIMESTAMP) \> CAST(DATE_FORMAT(CAST(`t_employeeattendance`.`CLOCK_IN_TIME` AS TIMESTAMP), 'yyyy-MM-dd') \|\| ' 09:00:00' AS TIMESTAMP) THEN 1 END) AS `late_count` FROM `t_employeeinfo` JOIN `t_employeeattendance` ON `t_employeeinfo`.`EMPLOYEE_ID` = `t_employeeattendance`.`EMPLOYEE_ID` GROUP BY `t_employeeinfo`.`EMPLOYEE_ID`, `t_employeeinfo`.`FULL_NAME`"}, "request_id": "b44bf49f-e299-4c17-8e72-90b991325a6e", "status": "data", "error_msg": ""}**
完成表结构替换后，将所有源表都替换成同一个表，导致程序执行错误，返回无法回答问题：
2025-10-17T03:25:46.329Z INFO 47 --- genie-backend exe-pool-1 com.jd.genie.service.Nl2SqlService : b44bf49f-e299-4c17-8e72-90b991325a6e,b44bf49f-e299-4c17-8e72-90b991325a6e 执行sql:SELECT employee_info.EMPLOYEE_ID, employee_info.FULL_NAME, COUNT(CASE WHEN CAST(employee_info.CLOCK_IN_TIME AS TIMESTAMP) > CAST(DATE_FORMAT(CAST(employee_info.CLOCK_IN_TIME AS TIMESTAMP), 'yyyy-MM-dd') || ' 09:00:00' AS TIMESTAMP) THEN 1 END) AS late_count FROM employee_info JOIN employee_info ON employee_info.EMPLOYEE_ID = employee_info.EMPLOYEE_ID GROUP BY employee_info.EMPLOYEE_ID, employee_info.FULL_NAME
原因分析：com.jd.genie.service.Nl2SqlService，代码中获取的TableName没有更新，导致被替换成同一张表
修复后，在循环替换过程中，更新tableName
修复效果：
2025-10-17T08:31:30.005Z INFO 46 --- [genie-backend] [48.237:1601/...] com.jd.genie.service.Nl2SqlService : 86f76794-e83c-42b3-9958-53904409e6f0,86f76794-e83c-42b3-9958-53904409e6f0 SSE数据结果：{"code": 200, "data": [{"query": "获取考勤打卡迟到员工的姓名和部门", "nl2sql": "SELECT DISTINCT `t_employeeinfo`.`FULL_NAME`, `t_employeeinfo`.`DEPARTMENT` FROM `t_employeeinfo` INNER JOIN `t_employeeattendance` ON `t_employeeinfo`.`EMPLOYEE_ID` = `t_employeeattendance`.`EMPLOYEE_ID` WHERE PARSEDATETIME(`t_employeeattendance`.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss') > PARSEDATETIME(CONCAT(FORMATDATETIME(PARSEDATETIME(`t_employeeattendance`.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss'), 'yyyy-MM-dd'), ' 09:00:00'), 'yyyy-MM-dd HH:mm:ss')"}], "request_id": "86f76794-e83c-42b3-9958-53904409e6f0", "status": "data", "error_msg": ""}
2025-10-17T08:31:30.018Z INFO 46 --- [genie-backend] [48.237:1601/...] com.jd.genie.service.Nl2SqlService : 86f76794-e83c-42b3-9958-53904409e6f0,86f76794-e83c-42b3-9958-53904409e6f0 SSE 连接关闭
2025-10-17T08:31:30.025Z INFO 46 --- [genie-backend] [ exe-pool-2] com.jd.genie.service.Nl2SqlService : 86f76794-e83c-42b3-9958-53904409e6f0 sse event count:208
2025-10-17T08:31:30.090Z INFO 46 --- [genie-backend] [ exe-pool-2] com.jd.genie.service.Nl2SqlService : *** 开始替换modelname：共有model[5]个
2025-10-17T08:31:30.091Z INFO 46 --- [genie-backend] [ exe-pool-2] com.jd.genie.service.Nl2SqlService : *** 开始匹配model[t_employeeinfo=employee_info],匹配前SQL：SELECT DISTINCT `t_employeeinfo`.`FULL_NAME`, `t_employeeinfo`.`DEPARTMENT` FROM `t_employeeinfo` INNER JOIN `t_employeeattendance` ON `t_employeeinfo`.`EMPLOYEE_ID` = `t_employeeattendance`.`EMPLOYEE_ID` WHERE PARSEDATETIME(`t_employeeattendance`.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss') > PARSEDATETIME(CONCAT(FORMATDATETIME(PARSEDATETIME(`t_employeeattendance`.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss'), 'yyyy-MM-dd'), ' 09:00:00'), 'yyyy-MM-dd HH:mm:ss')
2025-10-17T08:31:30.092Z INFO 46 --- [genie-backend] [ exe-pool-2] com.jd.genie.service.Nl2SqlService : *** 完成匹配model[t_employeeinfo=employee_info],匹配前SQL：SELECT DISTINCT employee_info.`FULL_NAME`, employee_info.`DEPARTMENT` FROM employee_info INNER JOIN `t_employeeattendance` ON employee_info.`EMPLOYEE_ID` = `t_employeeattendance`.`EMPLOYEE_ID` WHERE PARSEDATETIME(`t_employeeattendance`.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss') > PARSEDATETIME(CONCAT(FORMATDATETIME(PARSEDATETIME(`t_employeeattendance`.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss'), 'yyyy-MM-dd'), ' 09:00:00'), 'yyyy-MM-dd HH:mm:ss')
2025-10-17T08:31:30.092Z INFO 46 --- [genie-backend] [ exe-pool-2] com.jd.genie.service.Nl2SqlService : *** 开始匹配model[t_salaryinfo=salary_info],匹配前SQL：SELECT DISTINCT employee_info.`FULL_NAME`, employee_info.`DEPARTMENT` FROM employee_info INNER JOIN `t_employeeattendance` ON employee_info.`EMPLOYEE_ID` = `t_employeeattendance`.`EMPLOYEE_ID` WHERE PARSEDATETIME(`t_employeeattendance`.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss') > PARSEDATETIME(CONCAT(FORMATDATETIME(PARSEDATETIME(`t_employeeattendance`.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss'), 'yyyy-MM-dd'), ' 09:00:00'), 'yyyy-MM-dd HH:mm:ss')
2025-10-17T08:31:30.092Z INFO 46 --- [genie-backend] [ exe-pool-2] com.jd.genie.service.Nl2SqlService : *** 完成匹配model[t_salaryinfo=salary_info],匹配前SQL：SELECT DISTINCT employee_info.`FULL_NAME`, employee_info.`DEPARTMENT` FROM employee_info INNER JOIN `t_employeeattendance` ON employee_info.`EMPLOYEE_ID` = `t_employeeattendance`.`EMPLOYEE_ID` WHERE PARSEDATETIME(`t_employeeattendance`.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss') > PARSEDATETIME(CONCAT(FORMATDATETIME(PARSEDATETIME(`t_employeeattendance`.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss'), 'yyyy-MM-dd'), ' 09:00:00'), 'yyyy-MM-dd HH:mm:ss')
2025-10-17T08:31:30.092Z INFO 46 --- [genie-backend] [ exe-pool-2] com.jd.genie.service.Nl2SqlService : *** 开始匹配model[t_uegarulwipfivhutcvyawaoex=(select `order_date` as `采购日期`,product_id as `商品ID`,product_name as `商品名称`,category as `商品类型`,sub_category as `商品子类型` , quantity as `采购量` , sales as `采购总价格` from sales_data) t],匹配前SQL：SELECT DISTINCT employee_info.`FULL_NAME`, employee_info.`DEPARTMENT` FROM employee_info INNER JOIN `t_employeeattendance` ON employee_info.`EMPLOYEE_ID` = `t_employeeattendance`.`EMPLOYEE_ID` WHERE PARSEDATETIME(`t_employeeattendance`.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss') > PARSEDATETIME(CONCAT(FORMATDATETIME(PARSEDATETIME(`t_employeeattendance`.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss'), 'yyyy-MM-dd'), ' 09:00:00'), 'yyyy-MM-dd HH:mm:ss')
2025-10-17T08:31:30.093Z INFO 46 --- [genie-backend] [ exe-pool-2] com.jd.genie.service.Nl2SqlService : *** 完成匹配model[t_uegarulwipfivhutcvyawaoex=(select `order_date` as `采购日期`,product_id as `商品ID`,product_name as `商品名称`,category as `商品类型`,sub_category as `商品子类型` , quantity as `采购量` , sales as `采购总价格` from sales_data) t],匹配前SQL：SELECT DISTINCT employee_info.`FULL_NAME`, employee_info.`DEPARTMENT` FROM employee_info INNER JOIN `t_employeeattendance` ON employee_info.`EMPLOYEE_ID` = `t_employeeattendance`.`EMPLOYEE_ID` WHERE PARSEDATETIME(`t_employeeattendance`.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss') > PARSEDATETIME(CONCAT(FORMATDATETIME(PARSEDATETIME(`t_employeeattendance`.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss'), 'yyyy-MM-dd'), ' 09:00:00'), 'yyyy-MM-dd HH:mm:ss')
2025-10-17T08:31:30.111Z INFO 46 --- [genie-backend] [ exe-pool-2] com.jd.genie.service.Nl2SqlService : *** 开始匹配model[t_employeeattendance=employee_attendance],匹配前SQL：SELECT DISTINCT employee_info.`FULL_NAME`, employee_info.`DEPARTMENT` FROM employee_info INNER JOIN `t_employeeattendance` ON employee_info.`EMPLOYEE_ID` = `t_employeeattendance`.`EMPLOYEE_ID` WHERE PARSEDATETIME(`t_employeeattendance`.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss') > PARSEDATETIME(CONCAT(FORMATDATETIME(PARSEDATETIME(`t_employeeattendance`.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss'), 'yyyy-MM-dd'), ' 09:00:00'), 'yyyy-MM-dd HH:mm:ss')
2025-10-17T08:31:30.112Z INFO 46 --- [genie-backend] [ exe-pool-2] com.jd.genie.service.Nl2SqlService : *** 完成匹配model[t_employeeattendance=employee_attendance],匹配前SQL：SELECT DISTINCT employee_info.`FULL_NAME`, employee_info.`DEPARTMENT` FROM employee_info INNER JOIN employee_attendance ON employee_info.`EMPLOYEE_ID` = employee_attendance.`EMPLOYEE_ID` WHERE PARSEDATETIME(employee_attendance.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss') > PARSEDATETIME(CONCAT(FORMATDATETIME(PARSEDATETIME(employee_attendance.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss'), 'yyyy-MM-dd'), ' 09:00:00'), 'yyyy-MM-dd HH:mm:ss')
2025-10-17T08:31:30.113Z INFO 46 --- [genie-backend] [ exe-pool-2] com.jd.genie.service.Nl2SqlService : *** 开始匹配model[t_qtpbgamccmrctthlurauclckq=sales_data],匹配前SQL：SELECT DISTINCT employee_info.`FULL_NAME`, employee_info.`DEPARTMENT` FROM employee_info INNER JOIN employee_attendance ON employee_info.`EMPLOYEE_ID` = employee_attendance.`EMPLOYEE_ID` WHERE PARSEDATETIME(employee_attendance.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss') > PARSEDATETIME(CONCAT(FORMATDATETIME(PARSEDATETIME(employee_attendance.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss'), 'yyyy-MM-dd'), ' 09:00:00'), 'yyyy-MM-dd HH:mm:ss')
2025-10-17T08:31:30.113Z INFO 46 --- [genie-backend] [ exe-pool-2] com.jd.genie.service.Nl2SqlService : *** 完成匹配model[t_qtpbgamccmrctthlurauclckq=sales_data],匹配前SQL：SELECT DISTINCT employee_info.`FULL_NAME`, employee_info.`DEPARTMENT` FROM employee_info INNER JOIN employee_attendance ON employee_info.`EMPLOYEE_ID` = employee_attendance.`EMPLOYEE_ID` WHERE PARSEDATETIME(employee_attendance.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss') > PARSEDATETIME(CONCAT(FORMATDATETIME(PARSEDATETIME(employee_attendance.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss'), 'yyyy-MM-dd'), ' 09:00:00'), 'yyyy-MM-dd HH:mm:ss')
2025-10-17T08:31:30.113Z INFO 46 --- [genie-backend] [ exe-pool-2] com.jd.genie.service.Nl2SqlService : 86f76794-e83c-42b3-9958-53904409e6f0,86f76794-e83c-42b3-9958-53904409e6f0 执行sql:SELECT DISTINCT employee_info.`FULL_NAME`, employee_info.`DEPARTMENT` FROM employee_info INNER JOIN employee_attendance ON employee_info.`EMPLOYEE_ID` = employee_attendance.`EMPLOYEE_ID` WHERE PARSEDATETIME(employee_attendance.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss') > PARSEDATETIME(CONCAT(FORMATDATETIME(PARSEDATETIME(employee_attendance.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss'), 'yyyy-MM-dd'), ' 09:00:00'), 'yyyy-MM-dd HH:mm:ss')
2025-10-17T08:31:30.113Z INFO 46 --- [genie-backend] [ exe-pool-2] c.j.g.d.j.c.JdbcConnectionPools : 从缓存获取连接池 poolId genie-datasource
2025-10-17T08:31:30.113Z INFO 46 --- [genie-backend] [ exe-pool-2] c.j.g.d.provider.jdbc.JdbcDataProvider : jdbc执行sql:SELECT DISTINCT employee_info.`FULL_NAME`, employee_info.`DEPARTMENT` FROM employee_info INNER JOIN employee_attendance ON employee_info.`EMPLOYEE_ID` = employee_attendance.`EMPLOYEE_ID` WHERE PARSEDATETIME(employee_attendance.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss') > PARSEDATETIME(CONCAT(FORMATDATETIME(PARSEDATETIME(employee_attendance.`CLOCK_IN_TIME`, 'yyyy-MM-dd HH:mm:ss'), 'yyyy-MM-dd'), ' 09:00:00'), 'yyyy-MM-dd HH:mm:ss')
2025-10-17T08:31:30.114Z INFO 46 --- [genie-backend] [ exe-pool-2] c.j.g.d.j.connection.ConnectionWrapper : 获取数据库链接成功 poolId:genie-datasource
2025-10-17T08:31:30.373Z INFO 46 --- [genie-backend] [ exe-pool-2] com.jd.genie.service.Nl2SqlService : 86f76794-e83c-42b3-9958-53904409e6f0,86f76794-e83c-42b3-9958-53904409e6f0 查询sql结果大小：5