AXEtract

Low-Cost Cross-Domain Web Structured Information Extraction

---

AXEtract is a high-performance, low-cost framework for extracting structured data from web pages. Based on the paper "AXE: Low-Cost Cross-Domain Web Structured Information Extraction", it optimizes the extraction pipeline by using specialized LoRA adapters for pruning and query-specific extraction, enabling state-of-the-art results with small models (e.g., Qwen3-0.6B).

🚀 Key Features

• 🎯 Specialized LoRA Adapters: Uses task-specific adapters for DOM pruning and structured extraction, achieving high accuracy with minimal token overhead.
• ✂️ Smart DOM Pruning: Classifies and prunes irrelevant HTML nodes before passing them to the extractor, significantly reducing context window usage and costs.
• 📍 Grounded XPath Resolution (GXR): Automatically maps extracted JSON fields back to their original source XPaths in the DOM for verification and grounding.
• ⚡ High-Throughput Pipeline: Built-in support for multiple LLM engines, including vLLM for production-grade serving and HuggingFace for local research.
• 🌐 Cross-Domain Versatility: Designed to generalize across various web domains (e-commerce, real estate, listings) without needing domain-specific rules.

🛠️ Architecture

AXEtract follows a three-part decoupled pipeline for maximum efficiency:

1. Preprocessor: Fetches raw HTML and chunks it into manageable, token-aware fragments.
2. AI Extractor: Divided into two stages:
   • Pruner: A lightweight LLM (LoRA-powered) filters out noise and selects only relevant HTML chunks.
   • Extractor: A task-specific LLM maps the pruned HTML content directly to a structured JSON schema or natural language answer.
3. Postprocessor: Validates the output and resolves source XPaths via Grounded XPath Resolution (GXR).

📦 Installation

# Install from PyPI
uv pip install axetract

# Or install from source
git clone https://github.com/abdo-Mansour/axetract.git
cd axetract
uv sync

🚥 Quick Start

from pydantic import BaseModel
from axetract.pipeline import AXEPipeline

# 1. Initialize the pipeline with default LoRA adapters
# (Automatically downloads adapters from HuggingFace Hub)
pipeline = AXEPipeline.from_config(use_vllm=False)

# 2. Define your desired extraction schema
class Product(BaseModel):
    name: str
    price: str
    rating: float

# 3. Extract from a URL or raw HTML
url = "https://example.com/item/12345"
result = pipeline.extract(url, schema=Product)

# 4. Access your structured data
print(f"Status: {result.status}")
print(f"Prediction: {result.prediction}")
print(f"Source XPaths: {result.xpaths}")

🌐 API Server

AXEtract includes a built-in FastAPI server for high-throughput serving. After installing the package, start it with the installed CLI entry point:

axe-server

Or via python -m for development installs:

python -m axetract.server

Configuration is done via environment variables:

Variable	Default	Description
AXE_USE_VLLM	false	Set to true to use vLLM backend
AXE_PORT	8000	Port to listen on
AXE_HOST	0.0.0.0	Host to bind to
AXE_LOG_FILE	(stderr)	Optional path to a log file

See axe_server/client_example.py for examples of interacting with the API via requests.

📝 Citation

If you use AXEtract in your research, please cite our paper:

@misc{mansour2026axe,
      title={AXE: Low-Cost Cross-Domain Web Structured Information Extraction}, 
      author={Abdelrahman Mansour and Khaled W. Alshaer and Moataz Elsaban},
      year={2026},
      eprint={2602.01838},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2602.01838}, 
}

📜 License

This project is licensed under the MIT License - see the LICENSE file for details.