~95% on SimpleQA (e.g. Qwen3.6-27B on a 3090). Supports all local and cloud LLMs (llama.cpp, Ollama, Google, ...). 10+ search engines - arXiv, PubMed, your private documents. Everything Local & Encrypted.
Python academiaanthropicarxivbravedeep-researchencryptionhome-automationhomeserverlocallocal-deep-researchlocal-llmmistralollamaopenaipubmedresearchresearch-toolretrieval-augmented-generationsearxngself-hosted
"""Custom loader for legacy ``.xls`` (BIFF) spreadsheets.
``unstructured``'s Excel partitioner runs a ``msoffcrypto`` "is encrypted?"
pre-check on every workbook. That check crashes on some legacy ``.xls`` files
(``struct.error: unpack requires a buffer of 4 bytes`` in its xls97 record
reader) even though ``xlrd`` reads the same file fine. To make ``.xls`` uploads
extract reliably, this loader reads the workbook directly with pandas + xlrd
and bypasses that pre-check. Modern ``.xlsx`` files still use the unstructured
loader, whose OOXML path is unaffected.
"""
from pathlib import Path
from typing import Iterator
import pandas as pd
from langchain_core.document_loaders import BaseLoader
from langchain_core.documents import Document
from loguru import logger
class XLSLoader(BaseLoader):
"""Load a legacy ``.xls`` workbook and convert each sheet to text.
Every sheet is read with the ``xlrd`` engine and rendered as one
space-joined line per row (skipping empty cells), which mirrors the plain
text that the unstructured Excel loader produces for ``.xlsx``.
"""
def __init__(self, file_path: str | Path, **kwargs: object):
self.file_path = Path(file_path)
def lazy_load(self) -> Iterator[Document]:
try:
# sheet_name=None reads all sheets; header=None keeps the first row
# as data so column headers are included in the extracted text.
sheets = pd.read_excel(
self.file_path,
sheet_name=None,
header=None,
engine="xlrd",
# pandas reads the literal strings "N/A", "NA", "n/a", "NULL",
# "None", "NaN" and "nan" as missing values, so a cell saying any
# of them would be dropped by the pd.notna filter below. "N/A" is
# how a person writes "not applicable"; the cell means something.
keep_default_na=False,
)
except Exception as exc:
msg = str(exc).lower()
if "encrypt" in msg or "password" in msg:
raise ValueError(
f"XLS file is encrypted and cannot be read: {self.file_path}"
) from exc
logger.exception(f"Error loading XLS file: {self.file_path}")
raise
for sheet_name, frame in sheets.items():
lines = []
for row in frame.itertuples(index=False, name=None):
# A blank cell now arrives as "" rather than NaN, and is
# skipped the same way.
cells = [
text
for value in row
if pd.notna(value) and (text := str(value)) != ""
]
if cells:
lines.append(" ".join(cells))
text = "\n".join(lines).strip()
if not text:
continue
yield Document(
page_content=text,
metadata={
"source": str(self.file_path),
"file_type": "xls",
"sheet_name": sheet_name,
},
)
def load(self) -> list[Document]:
"""Load all sheets from the XLS file."""
return list(self.lazy_load())