Question answering on a table

[avantikajain]

I want to run the table-question-answering task on my table residing in postgres. I have tried using pg.transform() function to acheive the same. My query:
SELECT
vendor_name,
pgml.transform(
task => '{"task": "table-question-answering"
}'::JSONB,
inputs => ARRAY['{"question":"How many distinct vendors are present",
"table":"vendor_name"}']
-- inputs => array[vendor_name]
)
FROM pgml.vendor_summary
WHERE vendor_name IS NOT NULL
LIMIT 10;

Traceback:
ERROR: Traceback (most recent call last):
File "", line 196, in transform
File "", line 168, in call
File "/var/lib/postgresml/pgml-venv/lib/python3.10/site-packages/transformers/pipelines/table_question_answering.py", line 348, in call
pipeline_inputs = self._args_parser(*args, **kwargs)
File "/var/lib/postgresml/pgml-venv/lib/python3.10/site-packages/transformers/pipelines/table_question_answering.py", line 56, in call
raise ValueError(
ValueError: Keyword argument table should be a list of dict, but is <generator object TableQuestionAnsweringArgumentHandler.call.. at 0x7f9f511fa3b0>

SQL state: XX000

Wanted to know the correct syntax of the sql query to do question answering on existing table using postgresml

[montanalow]

Table QA is currently limited by GPTs to tables of a relatively small number of rows/columns, and this current workflow is limited to non SQL table formats, e.g. ascii tables depending on the model downloaded from huggingface. 😢

There has been some research into SQL scale table QA: https://arxiv.org/abs/2107.07653

As PostgresML matures, I'd love for this to become a primary use case.