MLXPipeline

A Local-First Data Extraction Pipeline powered by Apple MLX

Overview

MLXPipeline is a Python package for extracting, chunking, and analyzing data from various sources, optimized specifically for Apple Silicon. Unlike cloud-based solutions, MLXPipeline performs all processing, including machine learning inference tasks, locally on your machine.

Built on Apple's MLX framework, it efficiently leverages Apple's M-series chips to provide:

• 🔒 Enhanced Privacy: All data stays on your device
• 💰 Cost Efficiency: No API fees or usage limits
• ⚡ Speed: Optimized for Apple Silicon (M1/M2/M3 chips)
• 🔄 Versatility: Handles text, PDFs, webpages, images, audio, video, and more

Features

Content Extraction

• Document Parsing: Extract text from PDFs, DOCX, PPTX, HTML, Markdown, JSON, CSV
• Web Scraping: Extract content from websites with clean-up of navigation, ads, etc.
• OCR: Extract text from images using Optical Character Recognition
• Audio Transcription: Convert audio files to text using MLX-powered Whisper
• Video Transcription: Extract and transcribe audio from video files

Content Processing

• Text Chunking: Split large documents into manageable chunks based on size
• Semantic Chunking: Create chunks based on semantic similarity using embeddings
• Structured Information Extraction: Extract specific information using local LLMs

Local-First Machine Learning

• MLX-Powered: Uses Apple's MLX framework for ML tasks
• Embeddings: Generate text embeddings for semantic processing
• Transcription: Local audio transcription using MLX Whisper models
• LLM Integration: Use local Large Language Models for content extraction and analysis

Requirements

• macOS on Apple Silicon (M1/M2/M3 or newer)
• Python 3.9+
• MLX compatible models (not included in the package)

Installation

pip install mlxpipeline

Model Setup

MLXPipeline requires pre-downloaded MLX-compatible models to function. You should download:

1. MLX LLM Model: For extraction and analysis (e.g., Llama-3-8B-MLX)
2. MLX Whisper Model: For audio transcription
3. MLX Embedding Model: For semantic chunking (e.g., all-MiniLM-L6-v2)

These models should be placed in the default model directory ~/.mlxpipeline/models or specified using the appropriate arguments.

You can find MLX-compatible models at:

• MLX Community Models
• Hugging Face (models with MLX support)

Usage

MLXPipeline can be used both as a command-line tool and as a Python library.

Command Line Usage

Scraping a webpage:

mlxpipeline --source https://example.com --output result.txt

Chunking a document:

mlxpipeline --source document.pdf --chunk-type text --chunk-size 1000 --output chunks.txt

Extracting structured information:

mlxpipeline --source article.txt --extract --schema schema.json --output result.json

Audio transcription:

mlxpipeline --source recording.mp3 --whisper-model-path ~/.mlxpipeline/models/whisper-small-mlx --output transcription.txt

Using a custom LLM for extraction:

mlxpipeline --source document.pdf --extract --schema schema.json --llm-model-path ~/models/llama-3-8b-mlx --output result.json

Python Library Usage

Basic document processing:

from mlxpipeline.scraper import scrape_pdf
from mlxpipeline.chunker import chunk_text
from mlxpipeline.extract import extract_from_chunk

# Extract text from PDF
text = scrape_pdf("document.pdf")

# Split into chunks
chunks = chunk_text(text, chunk_size=1000, chunk_overlap=100)

# Extract structured information
schema = {
    "title": "string",
    "author": "string",
    "key_points": "array"
}

for i, chunk in enumerate(chunks):
    result = extract_from_chunk(
        chunk, 
        schema=schema,
        llm_model_path="~/.mlxpipeline/models/llm/llama-3-8b-mlx"
    )
    print(f"Chunk {i+1} extraction: {result}")

Webpage scraping and analysis:

from mlxpipeline.scraper import scrape_webpage, ai_extract_webpage_content

# Simple webpage scraping
content = scrape_webpage("https://example.com/article")

# AI-powered extraction (focuses on main content)
main_content = ai_extract_webpage_content(
    "https://example.com/article",
    llm_model_path="~/.mlxpipeline/models/llm/llama-3-8b-mlx"
)

Audio transcription:

from mlxpipeline.scraper import scrape_audio

# Transcribe audio
transcription = scrape_audio(
    "interview.mp3",
    whisper_model_path="~/.mlxpipeline/models/whisper/whisper-small-mlx"
)

Differences from thepipe_api

MLXPipeline is a fork of thepipe_api that has been completely redesigned to be local-first and Apple Silicon optimized. Key differences include:

1. Local-Only Processing: All ML tasks run locally with no cloud API dependencies
2. MLX Integration: Uses Apple's MLX framework for optimized performance on Apple Silicon
3. Model Management: Users must download and provide MLX-compatible models
4. Simplified Architecture: Removed all API interactions and authentication requirements
5. Performance Focus: Optimized for M-series chips with unified memory architecture

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

MIT License - See LICENSE file for details