skill-tools

A Node.js skill/library for AI agents to search PDFs/images, download and validate files, convert PDFs to Markdown, and generate speech locally with Supertone Supertonic ONNX models.

Install

npx skills add gafapa/skill-tools

Or clone manually:

git clone https://github.com/gafapa/skill-tools.git
cd skill-tools
npm install

Features

• searchPDFs(query, limit): PDF search via Bing HTML scraping (no API key)
• searchImages(query, limit): Google Images via google-img-scrap (no API key)
• downloadFile(url, outputDir, expectedType): streaming download + MIME validation + extension fix
• convertPdfToMarkdown(pdfPath, options): PDF to structured Markdown (unpdf / PDF.js)
• generateSpeech(text, outputPath, options): local TTS with ONNX Runtime + automatic model download

Usage

import {
  searchPDFs,
  searchImages,
  downloadFile,
  convertPdfToMarkdown,
  generateSpeech,
} from './index.js';
import fs from 'node:fs/promises';

Search

const pdfs = await searchPDFs('machine learning tutorial', 5);
const images = await searchImages('space wallpaper', 3);

Download

downloadFile() expects the destination directory to exist:

await fs.mkdir('./downloads', { recursive: true });
const savedPath = await downloadFile(pdfs[0].url, './downloads', 'pdf');

PDF to Markdown

const markdown = await convertPdfToMarkdown(savedPath, {
  detectHeadings: true,
  joinParagraphs: true,
  normaliseBullets: true,
  fixBrokenUrls: true,
  includeMetadata: false,
});

Text to Speech (automatic asset download)

On first run, the library downloads the required Supertonic 2 assets from Hugging Face into:

• ./models/supertonic-2/onnx/ (multiple ONNX files + JSON metadata)
• ./models/supertonic-2/voice_styles/ (selected voice style JSON)

await generateSpeech('Hello world!', './audio.wav', {
  voice: 'F1',   // F1-F5 or M1-M5, or a local .json voice-style path
  lang: 'en',    // en | ko | es | pt | fr
  speed: 1.0,
  steps: 20,
});

Full pipeline example

await fs.mkdir('./downloads', { recursive: true });

const [first] = await searchPDFs('javascript guide', 3);
const pdfPath = await downloadFile(first.url, './downloads', 'pdf');
const markdown = await convertPdfToMarkdown(pdfPath, { includeMetadata: true });

await fs.writeFile(pdfPath.replace('.pdf', '.md'), markdown);
await generateSpeech(markdown.slice(0, 400), './summary.wav', {
  voice: 'F1',
  lang: 'en',
  steps: 5,
});

Testing

Fast unit tests (default):

npm test

Integration tests are separated:

npm run test:integration:search   # set RUN_NETWORK_TESTS=1 to actually run
npm run test:integration:tts      # set RUN_TTS_INTEGRATION=1 to actually run

PowerShell examples:

$env:RUN_NETWORK_TESTS='1'; npm run test:integration:search
$env:RUN_TTS_INTEGRATION='1'; npm run test:integration:tts

Notes

• Search functions return [] on recoverable failures.
• Download / PDF conversion / TTS return null on recoverable failures.
• TTS inference runs locally after the one-time Supertonic 2 asset download.

License

MIT