Library of the Week — Texify

Texify — LaTeX extraction from mathematical documents, instantly

GitHub · Language: Python · License: Apache 2.0

What it does

Texify converts images and PDFs containing math — equations, formulas, mixed text — into clean LaTeX or Markdown with rendered math. It’s built for developers who need to ingest technical documents into LLM pipelines without losing mathematical structure, where OCR tools turn equations into garbage.

Why it stands out

• Purpose-built for math: General OCR (Tesseract, cloud vision APIs) fails badly on notation. Texify is fine-tuned specifically on mathematical content and handles multi-line equations, fractions, integrals, and Greek symbols that break generic tools
• Two output modes: Returns either raw LaTeX ($$\frac{d}{dx}...$$) or Markdown with inline math delimiters, making it easy to feed downstream into a vector store or renderer without post-processing
• Runs locally: Weights ship with the package — no API key, no rate limits, no data leaving your machine. Critical for academic and enterprise pipelines where math-heavy documents are often sensitive
• Pairs naturally with Marker: From the same author as Marker (PDF→Markdown), Texify handles the math layer that Marker delegates; the two compose cleanly for full document ingestion

Quick start

from texify.inference import batch_inference
from texify.model.model import load_model
from texify.model.processor import load_processor
from PIL import Image

model = load_model()
processor = load_processor()

img = Image.open("equation.png")
results = batch_inference([img], model, processor)

print(results[0])
# Output: $$\int_{0}^{\infty} e^{-x^2} dx = \frac{\sqrt{\pi}}{2}$$

When to use it

• RAG over technical documents: Physics papers, engineering specs, or financial models where equations are load-bearing — Texify preserves them instead of corrupting them into token soup
• Dataset construction: Building math instruction-tuning datasets from existing textbooks or papers without manual transcription
• Document QA pipelines: When your users will ask questions that require reasoning over formulas and your chunking strategy needs LaTeX that an LLM can actually parse

When to skip it

• Pure prose documents: If your PDFs are text-heavy with occasional simple fractions, the overhead isn’t justified — Marker alone or a basic PDF parser is faster and lighter
• Production throughput at scale: Texify runs on a vision encoder-decoder and is CPU-accessible but slow without a GPU; for high-volume batch jobs you’ll want to plan for inference infrastructure or pre-process offline

The verdict

Texify solves a genuinely painful gap in document ingestion pipelines that most tools paper over or ignore entirely. If you’re building RAG or fine-tuning datasets on STEM content, it belongs in your stack alongside Marker. The local-first design and clean Python API make integration straightforward — no excuses for mangling an integral sign into (d/dx) anymore.