llmaix 0.0.21

Add your description here

LLMAIxv2 Library

LLMAIx is an end‑to‑end toolkit for turning raw documents into structured knowledge with large‑language models. It now features a modular, Pydantic‑validated preprocessing core, richer OCR choices, and a revamped CLI.

Status: the public API is still stabilising; expect small breaking changes before a 2.0 stable release.

✨ Key capabilities

• Robust preprocessing –extract Markdown or plain text from PDFs, DOCX, TXT and images. The pipeline tries cheap text extraction first (PyMuPDF) and falls back to OCR (OCR‑my‑PDF/PaddleOCR/Surya) only when needed).
• Layout‑aware enrichment – advanced mode plugs into the Docling pipeline for tables, formulas and picture descriptions, optionally powered by a local or remote vision‑language model.
• MIME‑aware loading – files (or byte buffers) are classified with python‑magic so even extension‑less uploads are handled correctly.
• Information extraction – send arbitrary prompts + a Pydantic schema and get back valid JSON, using any OpenAI‑compatible LLM endpoint.
• CLI utilities – one‑command document conversion (llmaix preprocess) and structured‑info extraction (llmaix extract).
• Extensible – register new back‑ends (e.g. EPUB) with a single decorator; models and OCR engines can be swapped freely.

🛠 Installation

pip install llmaix          # base
pip install llmaix[docling] # + Docling/VLM extras
pip install llmaix[surya]   # + Surya‑OCR
pip install llmaix[all]     # everything

If you need GPU PaddleOCR:

uv pip install \
  --index-url https://www.paddlepaddle.org.cn/packages/stable/cu129/ \
  paddlepaddle-gpu==3.1.0

PaddleOCR supports 80+ languages out‑of‑the‑box). For MIME detection install libmagic (Linux/macOS) or python-magic-win64 on Windows).

🚀 Quick start

CLI

llmaix preprocess myscan.pdf                 # fast mode, auto‑OCR
llmaix preprocess doc.pdf --mode advanced \
    --enable-picture-description             # Docling + VLM captions
llmaix preprocess scan.pdf --force-ocr \
    --ocr-engine paddleocr -o out.md
llmaix extract -i "Acme Inc. raised $10 M..." # JSON extraction

Python API

from llmaix.preprocess import DocumentPreprocessor

# 1) simple PDF (born digital)
text = DocumentPreprocessor(mode="fast").process("report.pdf")

# 2) scanned PDF with multilingual OCR
proc = DocumentPreprocessor(
    mode="advanced",
    ocr_engine="surya",
    force_ocr=True,
    enable_picture_description=True,
    use_local_vlm=True,
    local_vlm_repo_id="HuggingFaceTB/SmolVLM-256M-Instruct",
    vlm_prompt="Please describe this document in detail.",
)
markdown = proc.process("scan_no_text.pdf")

Information extraction

from llmaix import extract_info
from pydantic import BaseModel

class LabInfo(BaseModel):
    name: str
    location: str
    lead: str

sentence = (
    "The KatherLab is a research group at TU Dresden led by Prof. Jakob N. Kather."
)
json_out = extract_info(
    prompt=f"Extract lab facts: {sentence}",
    pydantic_model=LabInfo,
    llm_model="gpt-4o-mini"
)
print(json_out.json(indent=2))

⚙️ Back‑end matrix

Task	Engine	Notes
Text extraction	PyMuPDF‑for‑LLM	Fast Markdown conversion from PDFs
	Docling	Layout‑aware; optional VLM captions
OCR	OCR‑my‑PDF (Tesseract)	Strong PDF/A support
	Surya‑OCR	Local transformer OCR, 90 + langs
	PaddleOCR PP‑Structure	Table & formula detection
MIME sniffing	python‑magic	libmagic signatures
(optional)	filetype	pure‑Python fallback

🧪 Tests

Clone and run:

git clone https://github.com/KatherLab/LLMAIx-v2.git
cd LLMAIx-v2
uv sync
uv run pytest          # full suite

You can focus on a backend:

uv run pytest tests/test_preprocess.py -k paddleocr