Table based output schema validation

[gstomar]

Hi All,

Trust you are well !

We are getting LLM output in table structure which we want to validate via Guardrails but I couldn't find any documents around it. Not sure whether we can do it or not. Can you please let me know if it's possible via any way ?

For e.g.

Quarter	Revenue (in billions)	Net Income (in billions)	Diluted EPS
Q1	43.1	15.2	6.41
Q2	46.4	15.9	6.67
Q3	44.9	12.9	5.33
Q4	51.7	14.3	5.65
FY	182.6	54.2	21.39

This is we need to validate. Just need to make sure that the data is being returned in Table format from LLM.
Is there any way or pointer which might help ?

Many Thanks for your time.

Kind Regards
Gaurav

[smohiuddin]

Hi Gauruv, can you share the file format you're using? Is this a Markdown table?

[gstomar]

Hi @smohiuddin
Yes, that's correct.
We are providing the context to LLM to return the output in table like the one I mentioned in my question.
But the tables are of not fixed column sizes. They will be changed based on input data and output configuration.
There is no fixed file format. It's all gets dynamically generated based on output schema config to LLM.
Please let me know if something is not clear.
Thanks

[zsimjee]

Hi Guarav!

You can write a simple validator that validates the markdown table format. Inherit the BaseValidator object from guardrails.validators and write the validator function to ensure that it's formatted correctly. See the regex validator for example

def validate():
lines = table_string.splitlines()

Check for header row

header_row = lines[0]
num_columns = len(re.findall(r"|", header_row)) + 1
if not header_row.startswith("|") or not header_row.endswith("|"):
return FailResult

Check for divider row

if not lines[1].startswith("---") or not lines[1].endswith("---"):
return FailResult

Check subsequent rows

for line in lines[2:]:
if not line.startswith("|") or not line.endswith("|"):
return FailResult
cells = re.findall(r"|([^\|]+)|", line)
if len(cells) != num_columns:
return FailResult

return PassResult

[gstomar]

Hi @zsimjee
Thanks for the detailed reply.
Much appreciated.
Will try it today.
Cheers
GT

[zsimjee]

Hey hey, curious to see if this worked! Were you able to try it?