fix(CSVDocumentSplitter): report column positions when read_csv_kwargs sets a header

[sclfcz]

Fixes #12965

What

CSVDocumentSplitter forces header=None so that columns are integer positions, and then reads col_idx_start straight off the column label:

resolved_read_csv_kwargs = {"header": None, "skip_blank_lines": False, "dtype": object, **self.read_csv_kwargs}
...
"col_idx_start": int(split_df.columns[0]),

A caller who passes read_csv_kwargs={"header": 0} (documented as supported, and "infer" behaves the same) gets the first row as string labels, so int("name") raised ValueError for every sub-table and the document could not be split at all:

splitter = CSVDocumentSplitter(row_split_threshold=1, column_split_threshold=None, read_csv_kwargs={"header": 0})
splitter.run([Document(content="name,score\nAda,9\n\nBob,8")])
# ValueError: invalid literal for int() with base 10: 'name'

Change

Map the labels back to their position in the original frame, so col_idx_start keeps meaning "starting column index of the sub-table in the original table" regardless of how the CSV was read:

column_positions = {label: position for position, label in enumerate(df.columns)}
...
"col_idx_start": column_positions[split_df.columns[0]],

Testing

read_csv_kwargs	before	after
default (header=None)	[(0, 0), (3, 0)]	[(0, 0), (3, 0)] (unchanged)
{"header": 0}	ValueError: invalid literal for int() with base 10: 'name'	[(0, 0), (2, 0)]
{"header": "infer"}	same ValueError	[(0, 0), (2, 0)]

The row indices differ because with header=0 the header row is consumed, so the sub-tables really do start at rows 0 and 2; col_idx_start stays 0.

• pytest test/components/preprocessors/test_csv_document_splitter_header_kwargs.py → 2 passed
• reverting only the col_idx_start line → the header test fails and the default case still passes

AI assistance

Written with an AI coding assistant (Claude-based agent): it located the cast, ran the before/after table above and wrote the test. I reviewed the diff, the reproduction output and the reasoning, and I take responsibility for the change.

[vercel]

@sclfcz is attempting to deploy a commit to the deepset Team on Vercel.

A member of the Team first needs to authorize it.

[github-actions]

Hi @sclfcz, thanks for your interest in contributing to Haystack! 🙏

⛔ First-time contributors can have at most 1 open pull request in this repository until it has been approved, so this PR was closed automatically. Your open pull request #12963 is unaffected. Once it has been approved by a maintainer, you are welcome to open more PRs. Feel free to reopen this one at that point.

See the contributing guidelines for details.

This is an automated message to help us keep the review queue healthy.

[sclfcz]

Added a second commit to the same branch, because it is the same label-vs-position confusion in the same method:

The sort key was dataframe.columns[0] — an integer position only while header=None. With a caller-supplied header the sub-tables were ordered alphabetically, so header=0 on z,_,a\n1,,2\n3,,4 (the empty middle column splits it) returned the sub-table at position 2 with split_id=0 and the one at position 0 with split_id=1, i.e. the table's columns came back swapped. It now sorts on the mapped position, exactly like col_idx_start.

read_csv_kwargs	before	after
default header=None	[(0, 0, 0)]	unchanged
{"header": 0}	[(0, 2, 0), (0, 0, 1)]	[(0, 0, 0), (0, 2, 1)]

pytest test/components/preprocessors/test_csv_document_splitter_header_kwargs.py → 3 passed; reverting only the sort key fails the ordering test and leaves the other two passing.

[sclfcz]

Understood - staying within the one-open-PR rule for first-time contributors.

This PR carries a real fix (two commits, both the same label-vs-position confusion) and its tests pass locally:

• header=0 no longer raises ValueError: int('name') when computing col_idx_start;
• sub-tables are ordered by column position instead of header label, so z,_,a keeps column 0 before column 2;
• pytest test/components/preprocessors/test_csv_document_splitter_header_kwargs.py → 3 passed, and reverting either change fails exactly one test.

I will reopen it as soon as #12963 has been reviewed, or let me know if you would rather I fold it into that PR.