The wrenai-service encountered a "KeyError: 'table_contents'" error.

[tatuke]

Hi, everyone. I set up the AI API settings and started using the chat. When I send a simple request like 'How many columns are in the table user?', it shows the answer pretty well. But when I ask about 'How many rows are in the table user?', it doesn't work as expected.The wrenai-service encountered a "KeyError: 'table_contents'" error.

  File "/src/pipelines/retrieval/retrieval.py", line 314, in construct_retrieval_results
    reformated_json[table["table_name"]] = table["table_contents"]
                                           ~~~~~^^^^^^^^^^^^^^^^^^
KeyError: 'table_contents'

At the beginning, I wondered if there was no more memory left. So, I deployed the service on a machine with 64GB of RAM and 30GB of free memory remaining. However, the same error occurred. The table 'user' only has a total of 700 rows and is 14KB in size. So here I am, looking for some help and advice.

The AI service image version I pulled was 0.15.7.

[cyyeh]

Hi, everyone. I set up the AI API settings and started using the chat. When I send a simple request like 'How many columns are in the table user?', it shows the answer pretty well. But when I ask about 'How many rows are in the table user?', it doesn't work as expected.The wrenai-service encountered a "KeyError: 'table_contents'" error.

  File "/src/pipelines/retrieval/retrieval.py", line 314, in construct_retrieval_results
    reformated_json[table["table_name"]] = table["table_contents"]
                                           ~~~~~^^^^^^^^^^^^^^^^^^
KeyError: 'table_contents'

At the beginning, I wondered if there was no more memory left. So, I deployed the service on a machine with 64GB of RAM and 30GB of free memory remaining. However, the same error occurred. The table 'user' only has a total of 700 rows and is 14KB in size. So here I am, looking for some help and advice.

The AI service image version I pulled was 0.15.7.

@tatuke what llm are you using, I am thinking maybe the issue is due to that llm is not correctly generating json output

[tatuke]

Hi, everyone. I set up the AI API settings and started using the chat. When I send a simple request like 'How many columns are in the table user?', it shows the answer pretty well. But when I ask about 'How many rows are in the table user?', it doesn't work as expected.The wrenai-service encountered a "KeyError: 'table_contents'" error.

  File "/src/pipelines/retrieval/retrieval.py", line 314, in construct_retrieval_results
    reformated_json[table["table_name"]] = table["table_contents"]
                                           ~~~~~^^^^^^^^^^^^^^^^^^
KeyError: 'table_contents'

At the beginning, I wondered if there was no more memory left. So, I deployed the service on a machine with 64GB of RAM and 30GB of free memory remaining. However, the same error occurred. The table 'user' only has a total of 700 rows and is 14KB in size. So here I am, looking for some help and advice.
The AI service image version I pulled was 0.15.7.

@tatuke what llm are you using, I am thinking maybe the issue is due to that llm is not correctly generating json output

I am using qwen-max through the Aliyun api

These are the parameters shown on their official website

Which shows that the maximum output token of this LLM can reach 8192. Is this enough to meet the output requirements of my table at this level? Or should I try a larger output token?

[tatuke]

I have found, and can confirm, that this is an error caused by a difference in the return format of the LLM api. Waiting for a more generic LLM protocol or mapping? .x.