BookWyrm: Production-Ready Multimodal Extraction
BookWyrm is designed to solve the last mile problem of complex document processing, transforming your most difficult PDFs into clean, semantically rich data. We combine the state-of-the-art in document intelligence into a single, robust, and fully automated API. Works seamlessly with our PDF extraction and RAG pipeline capabilities.
The result? You stop writing fragile preprocessing code and start building agents that matter. Perfect for back office automation.
Model Category
BookWyrm’s Function
Benefit for Developers
Document Layout Models
Analyzes the visual structure (columns, headers, text blocks) to determine the logical reading order and hierarchy.
Eliminates "broken sentences" and polluted chunks, ensuring reliable RAG retrieval and reduced hallucinations.
Multimodal Vision Transformers
Accurately performs OCR on text-in-images, recognizes and transcribes complex elements like charts, tables, and handwritten annotations.
Guaranteed high-fidelity data that captures all document elements, even from poor-quality scans.
Proprietary Post-Processing LLMs
Cleans, normalizes, and converts extracted table data into clean, embeddable formats (e.g., JSON), and strips out noise (headers/footers).
Output is instantly ready for your vector database. No manual cleaning or template-based pre-processing required.
Document Layout Models
BookWyrm's Function
Analyzes the visual structure (columns, headers, text blocks) to determine the logical reading order and hierarchy.
Benefit for Developers
Eliminates "broken sentences" and polluted chunks, ensuring reliable RAG retrieval and reduced hallucinations.
Multimodal Vision Transformers
BookWyrm's Function
Accurately performs OCR on text-in-images, recognizes and transcribes complex elements like charts, tables, and handwritten annotations.
Benefit for Developers
Guaranteed high-fidelity data that captures all document elements, even from poor-quality scans.
Proprietary Post-Processing LLMs
BookWyrm's Function
Cleans, normalizes, and converts extracted table data into clean, embeddable formats (e.g., JSON), and strips out noise (headers/footers).
Benefit for Developers
Output is instantly ready for your vector database. No manual cleaning or template-based pre-processing required.
Quick Start: Extract Structured PDF Data
BookWyrm makes the hardest part of document processing the easiest part of your pipeline.
Use the single extract-pdf endpoint (via SDK or REST) to get structured, multimodal data ready for your next-generation LLMs.
from typing import BinaryIO
from bookwyrm.models import PDFPage, PDFTextElement
# Load PDF file using raw bytes (recommended)
with open("document.pdf", "rb") as f:
f: BinaryIO
pdf_bytes: bytes = f.read()
pages: List[PDFPage] = []
for response in client.stream_extract_pdf(
pdf_bytes=pdf_bytes,
filename="document.pdf"
):
if hasattr(response, 'page_data'):
pages.append(response.page_data)
elif hasattr(response, 'total_pages') and hasattr(response, 'type') and response.type == "metadata":
print(f"Starting extraction of {response.total_pages} pages")
print(f"Extracted {len(pages)} pages")
page: PDFPage
for page in pages:
print(f"Page {page.page_number}: {len(page.text_blocks)} text elements")
element: PDFTextElement
for element in page.text_blocks[:3]: # Show first 3 elements
print(f" - {element.text[:50]}...")
# pages is List[PDFPage] where each PDFPage has:
# - page_number: int (1-based)
# - text_blocks: List[PDFTextElement]
# - tables: List[dict] (placeholder)
# - images: List[dict] (placeholder)
#
# Each PDFPage has:
# - page_number: int (1-based)
# - text_blocks: List[PDFTextElement]
# - tables: List[dict] (placeholder)
# - images: List[dict] (placeholder)
#
# Each PDFTextElement has:
# - text: str (extracted text)
# - confidence: float (0.0-1.0 OCR confidence)
# - bbox: List[List[float]] (raw polygon coordinates)
# - coordinates: PDFBoundingBox (x1, y1, x2, y2 rectangle)
Remove workplace friction and let BookWyrm handle multimodal pre-processing pain.