mdextract 1.0.0

Universal document parser — converts PDF, DOCX, XLSX, CSV to Markdown strings for AI pipelines

mdextract

Universal document → Markdown parser for AI pipelines.

Converts PDF, DOCX, XLSX, and CSV files into clean Markdown strings with a single function call. Designed to be the extraction layer in RAG systems, LLM pipelines, and document processing workflows.

import mdextract

text = mdextract.parse_file("quarterly_report.pdf")
response = llm.chat(f"Summarise this:\n\n{text}")

---

Features

Format	Output
.pdf	Markdown with headings (detected by font size) and GFM tables
.docx	Markdown preserving Word heading styles (Heading 1–6, Title) and tables
.xlsx	One # Sheet Name section + GFM table per worksheet
.csv	Single GFM Markdown table

• Zero configuration — just point it at a file
• Returns a string — no temp files, no disk I/O required
• Layout-aware for PDFs — tables are detected and rendered separately from body text; headings are inferred from font size
• Scanned PDF / OCR support — image-only pages are automatically processed with Tesseract; supports any language and tessdata_best high-accuracy models
• AI-pipeline friendly — output is plain UTF-8 Markdown, ready for chunking, embedding, or prompt injection

---

Installation

pip install mdextract

Or with uv:

uv add mdextract

---

Quickstart

Functional API (recommended)

import mdextract

# Any supported format — auto-detected from extension
text: str = mdextract.parse_file("report.pdf")
text: str = mdextract.parse_file("data.xlsx")
text: str = mdextract.parse_file("table.csv")
text: str = mdextract.parse_file("document.docx")

# Scanned / image-only PDF — French
text: str = mdextract.parse_file("rapport.pdf", ocr_lang="fra")

# Scanned PDF — French + English mixed, using high-accuracy models
BEST = r"C:\Program Files\Tesseract-OCR\tessdata_best"
text: str = mdextract.parse_file("rapport.pdf", ocr_lang="fra", tessdata_dir=BEST)

Per-format helpers

from mdextract import parse_pdf, parse_docx, parse_csv, parse_xlsx

text = parse_pdf("report.pdf")
text = parse_docx("contract.docx")
text = parse_csv("users.csv")
text = parse_xlsx("financials.xlsx")

Class API

Useful when you want to reuse an instance or save output to disk:

from mdextract import mdextract

parser = mdextract()

# Returns Markdown string
text = parser.parse_file("report.pdf")

# Also write to disk
text = parser.parse_file("report.pdf", output="report.md")

# Inspect supported formats
print(parser.supported_extensions)
# ['.csv', '.docx', '.pdf', '.xlsx']

---

AI Pipeline Examples

RAG (Retrieval-Augmented Generation)

import mdextract
from your_vectorstore import embed_and_store

for file in Path("docs/").glob("**/*"):
    try:
        markdown = mdextract.parse_file(str(file))
        embed_and_store(source=str(file), content=markdown)
    except ValueError:
        pass  # unsupported format, skip

LLM document Q&A

import mdextract
import openai

context = mdextract.parse_file("annual_report.pdf")

response = openai.chat.completions.create(
    model="gpt-4o",
    messages=[
        {"role": "system", "content": "You are a financial analyst."},
        {"role": "user", "content": f"Answer based on this document:\n\n{context}\n\nQuestion: What was the net revenue?"},
    ],
)

Batch processing

import mdextract
from pathlib import Path

results = {}
for path in Path("uploads/").iterdir():
    try:
        results[path.name] = mdextract.parse_file(str(path))
    except (ValueError, FileNotFoundError) as e:
        results[path.name] = f"Error: {e}"

---

Format Notes

PDF

• Character-level extraction via pdfplumber
• Tables detected automatically using ruling lines; table cells excluded from body text stream
• Headings detected by font size relative to the dominant body font size
• Page separators inserted as --- with <!-- Page N --> comments
• Scanned pages (no embedded text) are automatically OCR'd via Tesseract — no extra code needed

OCR parameters

Parameter	Default	Description
ocr_lang	"eng"	Tesseract language code(s). Use "fra" for French, "eng+fra" for mixed.
tessdata_dir	None	Path to a custom tessdata folder. Point this at tessdata_best for higher accuracy.

import mdextract

# Standard English OCR (default)
text = mdextract.parse_file("scan.pdf")

# French OCR — standard models
text = mdextract.parse_file("rapport.pdf", ocr_lang="fra")

# French OCR — high-accuracy models (tessdata_best)
BEST = r"C:\Program Files\Tesseract-OCR\tessdata_best"
text = mdextract.parse_file("rapport.pdf", ocr_lang="fra", tessdata_dir=BEST)

See OCR Setup below for installation instructions.

DOCX

• Heading levels mapped from Word's built-in styles (Heading 1 → #, Title → #, etc.)
• List items detected via w:numPr XML nodes and rendered as - item
• Merged table cells are handled; content is joined with a space

XLSX

• Each worksheet becomes a top-level section: # Sheet Name
• Fully empty rows at the end of a sheet are stripped
• Cell values are coerced to strings; None cells become empty strings
• Multi-sheet workbooks produce multiple sections separated by ---

CSV

• First row treated as the header
• UTF-8 BOM handled automatically (utf-8-sig encoding)
• Short rows padded to match the column count of the widest row

---

Error Handling

import mdextract

try:
    text = mdextract.parse_file("report.pdf")
except FileNotFoundError:
    print("File does not exist")
except ValueError as e:
    print(e)  # "Unsupported file type '.xyz'. Supported: .csv, .docx, .pdf, .xlsx"

---

OCR Setup (Scanned PDFs)

For PDFs that contain scanned images instead of embedded text, mdextract automatically falls back to Tesseract OCR. Two extra packages and a Tesseract installation are required.

1 — Install Tesseract

Windows: Download and run the UB Mannheim installer: https://github.com/UB-Mannheim/tesseract/wiki

Default install path: C:\Program Files\Tesseract-OCR\tesseract.exe

macOS:

brew install tesseract

Linux (Debian/Ubuntu):

sudo apt install tesseract-ocr

2 — Install Python OCR dependencies

pip install pymupdf pytesseract

Or with extras (if published with OCR extras):

pip install "mdextract[ocr]"

3 — Download language models

Standard models (faster, smaller — ~5 MB each):

# Windows — save to Tesseract's tessdata folder
$dir = "C:\Program Files\Tesseract-OCR\tessdata"
Invoke-WebRequest -Uri "https://github.com/tesseract-ocr/tessdata/raw/main/fra.traineddata" -OutFile "$dir\fra.traineddata"
Invoke-WebRequest -Uri "https://github.com/tesseract-ocr/tessdata/raw/main/eng.traineddata" -OutFile "$dir\eng.traineddata"

Best models (higher accuracy, larger — ~20 MB each):

# Windows — save to a separate tessdata_best folder
$best = "C:\Program Files\Tesseract-OCR\tessdata_best"
New-Item -ItemType Directory -Path $best -Force
Invoke-WebRequest -Uri "https://github.com/tesseract-ocr/tessdata_best/raw/main/fra.traineddata" -OutFile "$best\fra.traineddata"
Invoke-WebRequest -Uri "https://github.com/tesseract-ocr/tessdata_best/raw/main/eng.traineddata" -OutFile "$best\eng.traineddata"

All available languages: https://github.com/tesseract-ocr/tessdata_best

4 — Verify

& "C:\Program Files\Tesseract-OCR\tesseract.exe" --list-langs
# Should include: eng, fra (and any others you downloaded)

Using tessdata_best in code

import mdextract

BEST = r"C:\Program Files\Tesseract-OCR\tessdata_best"

text = mdextract.parse_file(
    "rapport.pdf",
    ocr_lang="fra",          # language code
    tessdata_dir=BEST,       # point at tessdata_best folder
)

Note: Every language you pass in ocr_lang must have a matching .traineddata file in the tessdata_dir folder. For ocr_lang="fra+eng" you need both fra.traineddata and eng.traineddata.

---

Requirements

• Python ≥ 3.11
• pdfplumber — PDF extraction
• python-docx — DOCX parsing
• openpyxl — XLSX parsing

CSV parsing uses the Python standard library only.

Optional (scanned PDF / OCR support):

• pymupdf — PDF page rendering to image
• pytesseract — Tesseract OCR wrapper
• Tesseract OCR ≥ 5 installed on the system

---

License

MIT