📑 Down Craft

Node.js package to simplify the process of converting documents (PDF, DOCX, PPTX, and XLSX) into Markdown format. It uses tesseract.js, mammoth, pdf.js, and turndown to convert documents to Markdown format. For PDFs, it also provides an option to use vLLMs (Vision Large Language Models) for advanced OCR capabilities (using the OpenAI API).

Online Web Demo

https://down-craft.dyndns.org/

Installation

npm install down-craft

Usage

import { downCraft } from 'down-craft';
import fs from 'fs/promises';

async function example() {
  // Read file buffer
  const fileBuffer = await fs.readFile('document.docx');
  
  // Convert to markdown (pass file buffer and file type)
  const markdown = await downCraft(fileBuffer, 'docx');
  
  console.log(markdown);
}

Supported File Types

• PDF (.pdf)
• Microsoft Word (.docx)
• Microsoft PowerPoint (.pptx)
• Microsoft Excel (.xlsx)

API

downCraft(fileBuffer, fileType?, options?)

Converts a document buffer to markdown format.

• fileBuffer (Buffer): The document buffer to convert
• fileType (string, optional): File type ('pdf', 'docx', 'pptx', 'xlsx'). If not provided the file type will be attempted to be auto-detected.
• options (Object, optional): Conversion options
  • pdfConverterType (string, optional): Converter to use for PDF files ('standard' | 'llm' | 'ocr'). Default: 'standard'
  • llmParams (Object, required for 'llm' converter): LLM configuration
    • baseURL (string): Base URL for the LLM API
    • apiKey (string): API key for the LLM service
    • model (string): Model to use for OCR
    • systemPrompt (string, optional): System prompt for the LLM (see .env.example for the default)
    • userPrompt (string, optional): User prompt for the LLM (see .env.example for the default)
    • temperature (number, optional): Temperature for the LLM (default: 0)

Returns: Promise - The markdown content

PDF Conversions

• Standard: Extracts text using standard techniques (images are ignored).
• vLLM: Uses a vLLM-based OCR model to extract text from PDFs (high fidelity, but much slower and requires an LLM API endpoint).
• OCR: Uses Tesseract.js for OCR (results are less accurate, but faster than using vLLM).

Special Features

vLLM-based PDF Conversion

For PDFs that require advanced OCR capabilities, you can use the vLLM converter:

const markdown = await downCraft(pdfBuffer, 'pdf', {
  pdfConverterType: 'llm',
  llmParams: {
    baseURL: 'https://api.llm-service.com',
    apiKey: 'your-api-key',
    model: 'your-model-name'
  }
});

This converter:

• Extracts embedded images from the PDF
• Converts PDF pages to high-quality images
• Uses vLLM-based OCR for accurate text extraction
• Automatically cleans up temporary files

The llmParams object will attempt to read environment variables for baseURL, apiKey, and model if you have them defined. See the .env.example file for an example (it also shows an example of how you can define your own user/system prompts), as well as various LLM providers / models.

License

This package is licensed under the Apache 2.0 license.
See LICENSE for details.