LLMFileAnalysisOperator drops CSV quoting, so values with commas look like extra columns #76386

Trying LLMFileAnalysisOperator on a small CSV on main (5114080, common.ai 1.1.0, macOS arm64, Python 3.12), the model gets the sample rows with their quotes removed. A value with a comma in it then looks like two values:

from pathlib import Path

from airflow.providers.common.ai.utils.file_analysis import build_file_analysis_request

Path("people.csv").write_text('name,city,amount\n"Smith, John","Vancouver, BC",120\nAlice,Toronto,80\n')
request = build_file_analysis_request(
    file_path="people.csv", file_conn_id=None, prompt="Total amount per city?",
    multi_modal=False, max_files=20, max_file_size_bytes=5 * 1024 * 1024,
    max_total_size_bytes=20 * 1024 * 1024, max_text_chars=100_000, sample_rows=10,
)
print(request.user_content)

Header: name, city, amount
Sample rows:
Smith, John, Vancouver, BC, 120
Alice, Toronto, 80

The header has 3 columns and the first row now shows 5 values, so a question like "total amount per city" has nothing reliable to work from. A quoted value with a line break in it is split across two lines the same way. This is the text execute() sends as the prompt:

_render_csv reads the file with csv.reader, which removes the quotes, then joins each row back with ", ":

JSON, Parquet and Avro samples go through dumps_masked, so they keep their structure (Parquet, for example: L543). CSV is the only one that loses it.

What I'd suggest: quote a value the way CSV does when it holds a comma, a quote or a line break, and leave the rest alone, so the current alpha, 1 format stays the same:

def _join_csv_values(values: list[str]) -> str:
    quoted = [
        '"' + value.replace('"', '""') + '"' if any(char in value for char in ',"\r\n') else value
        for value in values
    ]
    return ", ".join(quoted)

The first row then reads "Smith, John", "Vancouver, BC", 120.

I tried this locally:

• a new test in test_file_analysis.py (a quoted comma, an escaped quote and a line break) fails on main and passes with the change;
• test_file_analysis.py, test_llm_file_analysis.py and test_object_storage.py pass (120 passed, 8 skipped because fastavro isn't installed);
• running the operator itself with a FunctionModel that records its prompt shows the same thing before and after.

What should happen instead: each CSV value should reach the model intact, so the sample rows have the same columns as the header.

I didn't find an existing issue or PR for this. I have the fix and the test ready and will open a PR.

Drafted-by: Claude Code (Opus 5.5); reviewed by @aarmens702-hub before posting