Fix: Parse CSV Single-Turn Golden ToolCalls as JSON objects

[seankelley-dt]

Summary

This PR fixes #2564 by altering the parsing logic for loading goldens from CSV.

Context

Issue #2564 shows full repro code for this error.

Changes

• deepeval/dataset/dataset.py:EvaluationDataset.add_goldens_from_csv_file: Instead of simply splitting the row value string by the provided delimiter, parse the entire row value string using the same logic as EvaluationDataset.add_test_cases_from_csv_file since EvaluationDataset.save_as serialises the ToolCall objects in this way.

Tests

• Updated tests/test_core/test_datasets/goldens.csv to include example values for tools_called and expected_tools columns

[vercel]

@seankelley-dt is attempting to deploy a commit to the Confident AI Team on Vercel.

A member of the Team first needs to authorize it.

[seankelley-dt]

looks right — the old delimiter-based split couldn't handle serialized ToolCall objects. two things: (1) the CSV fixture update removes the old format entirely, but users who have existing CSV files with the old delimiter format will now get json.loads errors. might want a fallback that tries JSON first, then falls back to delimiter split for backward compat. (2) what happens when tools_called column contains an empty string vs "[]"? the get_column_data default changed from "" to "[]", but existing CSVs without the column would have "" from pandas fillna.

This is a great point @themavik. I've just updated the changes to fallback to the original functionality if the JSON decode fails.

On the point about the get_column_data default changing from "" to "[]", this will allow for the same functionality as before where, if the csv does not contain expected_tools_col_name, ["[]"] * len(df) is returned from get_column_data, the json decode succeeds and the subsequent list comprehension for tool in trimAndLoadJson(expected_tools_str) never runs.

Empty strings in the dataframe column won't be an issue as there's checks for empty strings before the JSON decode.

[seankelley-dt]

Actually on second thought, I have changed the get_column_data default back to "" to avoid lots of unnecessary JSON decode calls. If the column is missing, the empty string check will already skip any parsing.

[penguine-ip]

@seankelley-dt thank you!