mostlyright.core.formats.csv

mostlyright.core.formats.csv

CSV format — pandas-native serialization.

dumps emits standard CSV without the index column. loads parses via pd.read_csv with pandas’ default dtype inference. Useful for spreadsheet interchange and shell-level inspection, but lossy:

• All dtype information is lost on the wire; pandas re-infers on read. Int64 becomes int64 or float64 (depending on whether the column has nulls); Categorical becomes object (strings); boolean becomes object if it contained nulls.
• NaN and None and empty cells are all written as the empty string and all read back as NaN. The distinction between “missing numeric” and “missing string” is not preserved.
• Timezone-aware timestamps serialize as their ISO-8601 representation with offset, and pd.read_csv parses them back as object strings unless the caller explicitly asks for parse_dates.
• Floating-point round-trip is subject to CSV’s lossy decimal representation — pandas uses repr() formatting by default which is generally exact for IEEE-754, but precision-critical callers should prefer parquet.

For lossless transport prefer parquet. For LLM-friendly tabular transport prefer TOON.

Functions

dumps(df)	Serialize a DataFrame to a CSV string.
loads(data)	Parse a CSV string back into a DataFrame.

mostlyright.core.formats.csv.dumps(df)

Serialize a DataFrame to a CSV string.

Drops the index (index=False) to match the wire shape the catalog adapters emit. Column order is preserved.

Phase 6 W2-T6: accepts pandas OR polars input; polars frames are converted to pandas at the boundary so the wire bytes stay identical regardless of caller backend.

• Return type: str
• Parameters: df (DataFrame)

mostlyright.core.formats.csv.loads(data)

Parse a CSV string back into a DataFrame.

Uses pandas’ default dtype inference — see module docstring for the loss cases callers must be aware of.

• Return type: DataFrame
• Parameters: data (str)