Document

The Universal Document Parsing Gateway for Sayou Fabric.

sayou-document is a high-fidelity parsing engine that converts diverse document formats (PDF, DOCX, PPTX, XLSX, Images) into a unified, structured Document Object Model (DOM).

Unlike simple text extractors, it preserves the semantic structure of documents—headers, tables, charts, and layout coordinates—making it ideal for RAG applications that require layout awareness.

---

1. Architecture & Role

The Document engine acts as a normalizer. It accepts raw file bytes and applies the optimal Parser Strategy to output a structured SayouDocument.

graph LR
    File[Raw File] --> Pipeline[Document Pipeline]

    subgraph Parsers
        PDF[PDF Parser + OCR]
        Office[Office Parser]
        Img[Image Converter]
    end

    Pipeline -->|Type Detection| Parsers
    Parsers --> DOM[Structured DOM]

1.1. Core Features

• Smart Routing: Automatically detects file types (signatures) and selects the best parser.
• Hybrid Extraction: Combines native text extraction for digital PDFs with OCR fallback for scanned images.
• Strict Schema: Outputs a standardized hierarchy (Document > Page > Element) regardless of input format.

---

2. Supported Formats

sayou-document supports the following file types out-of-the-box.

Format	Strategy Key	Description
PDF	pdf	Extracts text, images, and TOC using PyMuPDF. Supports OCR.
Word	docx	Parses DOCX files, preserving heading levels and lists.
PowerPoint	pptx	Extracts text frames, speaker notes, and tables from slides.
Excel	xlsx	Converts sheets into table elements and extracts embedded charts.
Image	image	Auto-converts JPG/PNG/TIFF to PDF, then applies OCR.

---

3. Installation

Bash

pip install sayou-document

# For OCR support (requires Tesseract installed on OS)
pip install "sayou-document[ocr]"

---

4. Usage

The DocumentPipeline orchestrates file detection and parsing. It standardizes the input via the process method.

Case A: PDF Parsing (Standard)

Processes a PDF file to extract structured text and layout info.

Python

import os
from sayou.document import DocumentPipeline

file_path = "quarterly_report.pdf"
with open(file_path, "rb") as f:
    file_bytes = f.read()

doc = DocumentPipeline.process(
    data=file_bytes,
    metadata={"filename": os.path.basename(file_path)}
)

# 4. Result
print(f"File: {doc.file_name}, Pages: {doc.page_count}")
print(f"First Element: {doc.pages[0].elements[0].text}")

Case B: Office Documents (Word/Excel)

Parses Office formats while preserving table structures.

Python

from sayou.document import DocumentPipeline

with open("salary_table.xlsx", "rb") as f:
    file_bytes = f.read()

doc = DocumentPipeline.process(
    data=file_bytes,
    metadata={"filename": "salary_table.xlsx"}
)

# Access tables
tables = [e for p in doc.pages for e in p.elements if e.category == "table"]
print(f"Extracted {len(tables)} tables.")

Case C: Image with OCR

Automatically handles image conversion and OCR processing.

Python

from sayou.document import DocumentPipeline

# Initialize with OCR enabled
pipeline = DocumentPipeline(config={"use_ocr": True, "ocr_lang": "eng"})

with open("scanned_receipt.png", "rb") as f:
    file_bytes = f.read()

doc = pipeline.process(
    data=file_bytes,
    metadata={"filename": "scanned_receipt.png"}
)

print(f"OCR Result: {doc.pages[0].elements[0].text}")

---

5. Configuration Keys

Customize the parsing behavior via the config dictionary.

• use_ocr: (bool) Enable OCR for scanned pages or images.
• ocr_lang: (str) Tesseract language code (default: eng+kor).
• extract_images: (bool) Whether to extract embedded images to disk.
• table_strategy: (str) fast (text-based) or accurate (vision-based).

---

6. License

Apache 2.0 License © 2026 Sayouzone