fix: preserve row_number columns in CSVToDocument

docs-website/docs/pipeline-components/converters/csvtodocument.mdx

@@ -29,6 +29,9 @@ Converts CSV files to documents.
 The component uses UTF-8 encoding by default, but you may specify a different encoding if needed during initialization.
 You can optionally attach metadata to each document with a `meta` parameter when running the component.
 
+In row mode (`conversion_mode="row"`), `content_column` selects the column used as document content.
+The remaining columns become metadata, and `meta["row_number"]` records the zero-based row index.
+
 ## Usage
 
 ### On its own

haystack/components/converters/csv.py

@@ -215,7 +215,7 @@ def _build_document_from_row(
             Remaining row columns are added to ``meta`` with collision-safe
             keys (prefixed with ``csv_`` if needed).
         """
-        row_meta = dict(base_meta)
+        row_meta = {**base_meta, "row_number": row_index}
 
         # content (strict: content_column must exist; validated by caller)
         content = self._safe_value(row.get(content_column))
@@ -235,5 +235,4 @@ def _build_document_from_row(
                     suffix += 1
             row_meta[key_to_use] = self._safe_value(v)
 
-        row_meta["row_number"] = row_index
         return Document(content=content, meta=row_meta)

releasenotes/notes/preserve-csv-row-number-column-7965b5c7df176503.yaml

@@ -0,0 +1,6 @@
+---
+fixes:
+  - |
+    ``CSVToDocument`` in row mode now preserves values from a CSV column named
+    ``row_number`` as ``csv_row_number`` in metadata (with a suffix on collision),
+    instead of overwriting them with the generated row index.

test/components/converters/test_csv_to_document.py

@@ -4,6 +4,7 @@
 
 import logging
 import os
+from pathlib import Path
 
 import pytest
 
@@ -138,28 +139,38 @@ def test_row_mode_with_content_column(self, tmp_path):
         assert docs[0].meta["row_number"] == 0
         assert os.path.basename(f) == docs[0].meta["file_path"]
 
-    def test_row_mode_meta_collision_prefixed(self, tmp_path):
-        # ByteStream meta has file_path and encoding; CSV also has those columns.
-        csv_text = "file_path,encoding,comment\r\nrowpath.csv,latin1,ok\r\n"
-        f = tmp_path / "collide.csv"
-        f.write_text(csv_text, encoding="utf-8")
-        bs = ByteStream.from_file_path(f)
-        bs.meta["file_path"] = str(f)
-        bs.meta["encoding"] = "utf-8"
+    def test_row_mode_row_number_as_content_column(self) -> None:
+        source = ByteStream(data=b"row_number,author\nrecord-42,Ada\n")
+        converter = CSVToDocument(conversion_mode="row")
 
-        conv = CSVToDocument(conversion_mode="row")
-        out = conv.run(sources=[bs], content_column="comment")
-        d = out["documents"][0]
-        # Original meta preserved
-        assert d.meta["file_path"] == os.path.basename(str(f))
-        assert d.meta["encoding"] == "utf-8"
-        # CSV columns stored with csv_ prefix (no clobber)
-        assert d.meta["csv_file_path"] == "rowpath.csv"
-        assert d.meta["csv_encoding"] == "latin1"
-        # content column isn't duplicated in meta
-        assert "comment" not in d.meta
-        assert d.meta["row_number"] == 0
-        assert d.content == "ok"
+        documents = converter.run(sources=[source], content_column="row_number")["documents"]
+
+        assert len(documents) == 1
+        assert documents[0].content == "record-42"
+        assert documents[0].meta == {"author": "Ada", "row_number": 0}
+
+    @pytest.mark.parametrize("column_name", ["file_path", "row_number"])
+    def test_row_mode_meta_collision_prefixed(self, tmp_path: Path, column_name: str) -> None:
+        # file_path collides with source metadata; row_number collides with the generated row index.
+        csv_text = f"{column_name},encoding,comment\r\nsource-value,latin1,ok\r\n"
+        path = tmp_path / "collide.csv"
+        path.write_text(csv_text, encoding="utf-8")
+        source = ByteStream.from_file_path(path)
+        source.meta["file_path"] = str(path)
+        source.meta["encoding"] = "utf-8"
+        converter = CSVToDocument(conversion_mode="row")
+
+        documents = converter.run(sources=[source], content_column="comment")["documents"]
+
+        assert len(documents) == 1
+        assert documents[0].content == "ok"
+        assert documents[0].meta == {
+            "file_path": "collide.csv",
+            "encoding": "utf-8",
+            "row_number": 0,
+            f"csv_{column_name}": "source-value",
+            "csv_encoding": "latin1",
+        }
 
     def test_row_mode_meta_collision_multiple_suffixes(self, tmp_path):
         """