Skip to main content

Common Issues

Data Query Failure

Troubleshooting approach:

  1. Check whether SQL generation is wrong, including: whether SQL syntax is wrong, whether table names and field names in SQL are consistent with the dataset.
  2. Check "Topic Details-Basic Configuration", whether there are spaces and special symbols in dataset table names. For large models, spaces and special symbols are very difficult to understand and prone to generation errors. It's recommended to remove related spaces and special symbols here, then check whether SQL generation will go wrong.
  3. Check whether table names and field names have duplication. Currently, if table names and field names are consistent in query SQL, it will cause BI-side SQL query errors.
  4. Check interface errors, exclude SSO configuration issues.

  • Cause: Since ChatBI relies on SSO to get user cookies, when operations configure incorrect SSO data, cookie acquisition will fail, ultimately manifested in SQL execution as unable to execute

  • Solution

  • If using data_synapse version before 2.2.0, if the user doesn't have permissions for the corresponding dataset in the SQL, it will cause SQL query failure. Need to upgrade to version >=2.2.0 (in this version, as long as you have topic permissions, you can query), or authorize the user with related dataset permissions

  • Confirm that the SSO provided by the user is private_key, not public_key. If provided incorrectly, need to regenerate and insert SSO

  • Operations need to check whether the SSO encoded database data is inserted into the SSO table in pg. If not inserted, please re-insert

  • Operations need to check the generation themselves, whether the generated sso token data is standard, whether domain_id and user_id base64 decoded don't have spaces or line breaks (if using command line, spaces may appear). If there are issues, please fix before inserting

Don't want to display enterprise LOGO in Q&A entry

In the "Management Center > Enterprise Configuration > Enterprise Visual > LOGO and Appearance > LOGO and Name" interface, uncheck the "Display" button.

Cannot enter ChatBI backend

Please check in "Management Center > User Management > Roles" configuration whether the user has "ChatBI Edit" permissions.

Cannot see query entry/topics

  • For the problem of not being able to see the query entry, please check in "Management Center > User Management > Roles" configuration whether the user has "ChatBI View" permissions.
  • For the problem of not being able to see query topics, please check according to the following steps:
  1. Check whether the topic is enabled.
  2. Check whether the query user has user permissions for the current topic.

ChatBI backend missing permission management module

For the problem of not being able to see the permission management module, please check in "Management Center > User Management > Roles" configuration whether the user has "ChatBI Authorization" permissions.

Cannot ask questions

When the frontend reports an error "Current question balance insufficient, please contact administrator", please contact GuanYuan Data Customer Success Manager for recharge. All customer environments have a default quota of 5000 questions, and cooperative customers will have adjusted agreed quotas.

Accuracy Problem Troubleshooting

Except for product errors, we receive the following types of problems from users.

Data Query Errors

The data results replied by ChatBI are not what users want

  1. Understand the data structure and caliber definition that users want, clarify the error content

    1. Query returns no data: Prioritize checking whether SQL generation is correct (focus on whether table names and field names are consistent with dataset maintenance). If SQL query logic is correct, then follow this logic to filter and check in the dataset, see if there is data in the data source.

    2. Missing time range specification: If in user questions, the time definition is relatively vague (for example, the question is "How is recent sales"), need to check whether the meaning of "recent" has been defined in general knowledge. If there's no related knowledge definition, need to first supplement in business knowledge base-general knowledge.

    3. Aggregation dimension error, metric selection error: If in user questions, there's no clear dimension information (for example, specifying SKU, region, store, etc.), need to check whether there's related knowledge in business knowledge base, whether there are problem+SQL pairs in error set that define the current problem and answer.

      1. Business knowledge base example: When users ask about <...>, please associate organization dataset and daily report dataset according to organization code field, summarize all organizations with organization level value of 3, including sub-organizations, summarize all attendance hours by organization code and organization name
      2. Error set example: Question: 2023 March population asset distribution of a brand supermarket channel Query SQL: ```SELECT population_asset_type AS population_asset_type, COUNT(DISTINCT consumer_unique_identifier) AS person_count FROM member_table WHERE affiliated_channel = 'offline' AND brand = 'Xuesonglan' AND CAST('2023-03-31' AS DATE) BETWEEN TO_DATE (start_time) AND TO_DATE (end_time) GROUP BY `population_asset_type````
      3. Metric generation error: Check whether the metrics in the current reply have caliber definitions in business knowledge base/error set. If not, need to first define and supplement in business knowledge base/error set.

  2. View operation logs, check whether related knowledge was retrieved during the Q&A process. If knowledge has been maintained but not actually retrieved, please contact GuanYuan for processing.

    Note

    General knowledge is retrieved in every conversation and not displayed in operation logs.

  3. If the above steps cannot solve the problem, please contact GuanYuan for processing.

Visualization Generation Error

Specify to generate a certain chart type, but the generation result is a table/no visualization chart is generated, directly generates a table

  1. Please check the current BI version, whether it's version 7.0 or above. In version 6.6, the "specify visualization chart" capability doesn't work.
  2. Please check the current table data structure, whether it can be converted to BI visualization charts. For example, if the returned data structure is dimension-time, dimension-SKU, metric-sales amount, the chart itself cannot be rendered as a bar chart in BI, which is a normal phenomenon.
  3. If the above steps cannot solve the problem, please contact GuanYuan for processing.

Topic Usage Q&A

  1. Q: How to improve Q&A accuracy?

    1. Ensure questions conform to the basic structure of effective questions, time/conditions/metrics are clear and unambiguous, avoid ambiguity
    2. Ensure data quality, avoid dirty and poor data that increases knowledge maintenance difficulty
    3. According to the list of questions that need to be answered, add knowledge targeted (general knowledge, business knowledge, error sets)

  2. Q: Can adding questions in "Test" improve accuracy?

    No. The "Test" function is only for batch validation of Q&A effectiveness, used to evaluate Q&A accuracy

  3. Q: For complex similar knowledge, is it better to maintain item by item or maintain together? Some long knowledge can be learned, some cannot

    Business knowledge is recommended to maintain item by item, avoid putting unrelated business knowledge in the same item

  4. Q: Knowledge base writing expressions vary from person to person. To reduce others' reading comprehension costs, are there recommended unified formats? Such as Markdown syntax, etc.

    Knowledge base has no mandatory format requirements, similar to prompt descriptions, just meet the requirements of clear logic and clear expression. For example: "When questions involve xxx, default to query/display/calculate...", "When questions involve aaa, bbb, ccc, all refer to field1", "When questions involve aaa, bbb, ccc, all refer to the same concept: field2= xxx", "metric name = specific calculation formula"

  5. Q: After knowledge base content is rich, how to quickly locate knowledge that needs modification? How to identify conflicts between new knowledge and historical knowledge?

    Whether to add knowledge depends on Q&A effectiveness. For knowledge base add/delete/modify, it's recommended to prioritize testing Q&A effectiveness. When there are questions that cannot be satisfied (knowledge missing/wrong/redundant), naturally you can clarify whether it's needed and the corresponding add/delete/modify content. "Business Knowledge" supports search to quickly find knowledge containing keywords. Before adding new knowledge, you can search for reference first

  1. Q: How to fuzzy match and hit the correct field? For example: when querying fields with similar names, or fields with similar enumeration values, and cannot 100% enumerate (enumeration volume is huge), how to accurately select field values?

    1. First confirm whether enumeration value learning is completed (Basic Configuration-Dataset-Expand or check operation logs for Value examples)
    2. Second, prioritize letting the large model directly find corresponding values based on proper nouns contained in questions. If completely identical to enumeration values still needs to record bad cases, for not completely identical, it's recommended to guide the large model through fuzzy matching, refer to the business knowledge in Temporary Solutions above

  1. From major categories, medium categories, minor categories, series, sub-series, etc., to select the correct field, can only trigger counter-questions to let business choose the field they need? Are there other solution suggestions?

    1. If the nouns contained in questions have completely identical enumeration values in multiple fields, can only let users choose through counter-questions. For example, major region and city field enumeration values are all prefecture-level cities, need to add business knowledge: "When users mention , please let users clarify the city name in the question, specifically whether it refers to city or major region."
    2. If enumeration values of different fields are not completely identical, then it's recommended to clarify query standards, which proper nouns refer to field1, which refer to field2. For example: "When questions involve aaa, bbb, ccc, all refer to major category, when involving ddd, eee all refer to series."

How to Get Logs for GuanYuan Problem Troubleshooting

Users can download Data_Synapse service logs in the "Management Center > Operations Management > Operation Logs" module to provide to GuanYuan for problem troubleshooting. If users are private large model integration customers, they also need to download and provide Data_Mind service logs to GuanYuan.