unocr 🦜

Unified OCR library with multi-driver support for Tesseract.js and AI vision models, providing structured text extraction using hast-based output format.

✨ Features

• 🔍 Multi-Engine Support: Tesseract.js and AI vision drivers with unified interface
• 📝 Structured Output: Hast-based OCR results for rich document structure
• 🌐 Universal Input: Support for various image formats via undio integration
• ⚡️ High Performance: Parallel processing with scheduler support
• 🔄 Batch Processing: Efficient batch OCR operations with configurable parallelism
• 🛡️ TypeScript: Full TypeScript support with comprehensive type definitions
• 🎯 Driver Architecture: Extensible driver system for easy engine integration
• 🤖 AI-Powered: Advanced AI vision models for enhanced OCR accuracy
• 🔧 Flexible AI Configuration: Customizable system prompts and model parameters
• 📊 Rich Metadata: Comprehensive processing metadata and engine information
• 🔧 Flexible Options: Customizable OCR options for different use cases

📥 Installation

# Using npm
npm install unocr

# Using yarn
yarn add unocr

# Using pnpm
pnpm add unocr

🚀 Basic Usage

Tesseract.js Driver

import { createOCRManager } from "unocr";
import tesseractDriver from "unocr/drivers/tesseract";

// Create OCR manager with Tesseract driver
const ocr = createOCRManager({
  driver: tesseractDriver({
    langs: ["eng", "chi_sim"], // English and Chinese
    logger: (m) => console.log(m), // Progress logging
  }),
});

// Single image OCR
const result = await ocr.recognize(imageBuffer);
console.log(result); // hast Root object

// Batch OCR with parallel processing
const results = await ocr.recognizes(imageArray, { parallel: 4 });
console.log(results); // Array of hast Root objects

// Clean up
await ocr.dispose();

AI Vision Driver

import { createOCRManager } from "unocr";
import aiDriver from "unocr/drivers/ai";
import { createOpenAICompatible } from "@ai-sdk/openai-compatible";

// Create AI model client
const openai = createOpenAICompatible({
  name: "openai",
  baseURL:
    process.env.OPENAI_COMPATIBLE_URL ||
    "https://dashscope.aliyuncs.com/compatible-mode/v1",
  apiKey: process.env.OPENAI_API_KEY,
});

const modelName = process.env.AI_MODEL || "qwen3-vl-flash";

// Create OCR manager with AI driver
const ocr = createOCRManager({
  driver: aiDriver({
    model: openai(modelName),
    system:
      "Extract all text from this image and return it as HTML. Use appropriate tags like h1-h6 for headings, p for paragraphs, and ul/ol for lists.",
  }),
});

// Single image OCR with AI
const result = await ocr.recognize(imageBuffer);
console.log(result); // hast Root object

// Batch OCR with AI
const results = await ocr.recognizes(imageArray, { parallel: 2 });
console.log(results); // Array of hast Root objects

// Clean up
await ocr.dispose();

🔧 Advanced Usage

🎯 Custom Driver Configuration

Tesseract.js Advanced Configuration

import { createOCRManager } from "unocr";
import tesseractDriver from "unocr/drivers/tesseract";

// Advanced Tesseract configuration
const ocr = createOCRManager({
  driver: tesseractDriver({
    langs: ["eng", "fra", "deu"],
    oem: 1, // LSTM only
    corePath: "./tesseract-core",
    langPath: "./lang-data",
    cacheMethod: "write",
    logger: (progress) => {
      if (progress.status === "recognizing text") {
        console.log(`Progress: ${progress.progress * 100}%`);
      }
    },
  }),
});

const result = await ocr.recognize(image);

AI Vision Custom Configuration

import { createOCRManager } from "unocr";
import aiDriver from "unocr/drivers/ai";
import { createOpenAICompatible } from "@ai-sdk/openai-compatible";

const openai = createOpenAICompatible({
  name: "openai",
  baseURL:
    process.env.OPENAI_COMPATIBLE_URL ||
    "https://dashscope.aliyuncs.com/compatible-mode/v1",
  apiKey: process.env.OPENAI_API_KEY,
});

const modelName = process.env.AI_MODEL || "qwen3-vl-flash";

// Advanced AI configuration
const ocr = createOCRManager({
  driver: aiDriver({
    model: openai(modelName),
    system:
      "You are an expert OCR system. Extract all visible text with precise formatting and return it as structured HTML.",
    temperature: 0.1,
    maxOutputTokens: 4000,
    maxRetries: 3,
  }),
});

const result = await ocr.recognize(image);

📊 Batch Processing with Custom Parallelism

import { createOCRManager } from "unocr";
import tesseractDriver from "unocr/drivers/tesseract";

const ocr = createOCRManager({
  driver: tesseractDriver({ langs: "eng" }),
});

// Process many images efficiently
const images = [image1, image2, image3, image4, image5];

// Use 2 workers for lower resource usage
const results = await ocr.recognizes(images, { parallel: 2 });

// Use maximum parallelism (up to image count)
const maxResults = await ocr.recognizes(images, { parallel: images.length });

await ocr.dispose();

🌐 Input Format Support

import { createOCRManager } from "unocr";
import tesseractDriver from "unocr/drivers/tesseract";

const ocr = createOCRManager({
  driver: tesseractDriver({ langs: "eng" }),
});

// Various input formats supported via undio
const imageInputs = [
  "https://example.com/image.jpg", // URL (string)
  ArrayBuffer, // ArrayBufferLike
  Uint8Array.from([]), // Uint8Array
  "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAA", // Base64 data URL
  new Blob(), // Blob
  new ReadableStream(), // ReadableStream
  new Response(), // Response
];

const results = await ocr.recognizes(imageInputs);
await ocr.dispose();

🔍 Working with Hast Output

import { createOCRManager } from "unocr";
import tesseractDriver from "unocr/drivers/tesseract";
import { toHtml } from "hast-util-to-html";

const ocr = createOCRManager({
  driver: tesseractDriver({ langs: "eng" }),
});

const result = await ocr.recognize(image);

// Convert hast to HTML
const html = toHtml(result);
console.log(html);
// <div class="ocr_page">...</div>

// Extract text content
function extractText(node): string {
  if (node.type === "text") {
    return node.value;
  }
  if (node.children) {
    return node.children.map(extractText).join("");
  }
  return "";
}

const text = extractText(result);
console.log(text);
// "Extracted text content"

await ocr.dispose();

📚 API Reference

🔧 Manager Creation

createOCRManager(options: OCRManagerOptions)

Create an OCR manager with unified API for text recognition.

import { createOCRManager } from "unocr";
import tesseractDriver from "unocr/drivers/tesseract";

const ocr = createOCRManager({
  driver: tesseractDriver({
    langs: "eng",
    logger: console.log,
  }),
});

const result = await ocr.recognize(image);
const results = await ocr.recognizes(images, { parallel: 4 });
await ocr.dispose();

🖼️ Input and Output Types

OCRInput

Universal input type supporting various image formats via undio integration:

• string - URLs or base64 data URLs
• ArrayBufferLike - ArrayBuffer and similar types
• Uint8Array - Typed array data
• Blob - File/Blob objects
• ReadableStream - Stream data
• Response - Fetch API Response objects

OCRResult

Structured OCR output in hast format for rich document structure.

🚗 Available Drivers

tesseractDriver(options?: TesseractOptions)

Create a Tesseract.js-based OCR driver with advanced configuration options.

aiDriver(options?: AIDriverOptions)

Create an AI vision-based OCR driver with support for multimodal AI models like GPT-4 Vision.

AI Driver Options:

• model - AI model instance (from @ai-sdk/openai-compatible or other AI SDK providers)
• system - System prompt for text extraction instructions
• temperature - Response randomness (0-1)
• maxOutputTokens - Maximum tokens in AI response
• maxRetries - Number of retry attempts on failure
• Additional AI SDK parameters supported

⚡ Performance

📊 Benchmarks

• 🚀 Multi-Engine Support: Choose between traditional OCR and AI vision models
• ⚡️ Parallel Processing: Configurable parallel worker execution
• 📦 Efficient Memory: Worker reuse and proper cleanup
• 🔄 Batch Operations: Optimized batch processing with scheduler
• 🤖 AI Accuracy: Enhanced text recognition with AI vision models
• 🔧 Flexible Processing: Local Tesseract for speed, AI for complex layouts

🎯 Performance Tips

// Reuse OCR manager for multiple operations
const ocr = createOCRManager({
  driver: tesseractDriver({ langs: "eng" }),
});

// Batch process when possible
const results = await ocr.recognizes(images, { parallel: 4 });

// Configure appropriate parallelism based on hardware
const cpuCount = navigator.hardwareConcurrency || 4;
const results = await ocr.recognizes(images, { parallel: cpuCount });

// Choose driver based on use case
const fastOcr = createOCRManager({
  driver: tesseractDriver({ langs: "eng" }), // Fast for simple documents
});

const accurateOcr = createOCRManager({
  driver: aiDriver({ model: openai(modelName) }), // Better for complex layouts
});

await ocr.dispose();

🔧 Configuration

Batch Processing

Configure parallel processing for batch operations:

const ocr = createOCRManager({
  driver: tesseractDriver({ langs: "eng" }),
});

// Process with 2 workers (conservative)
await ocr.recognizes(images, { parallel: 2 });

// Process with 8 workers (high performance)
await ocr.recognizes(images, { parallel: 8 });

🌐 Ecosystem Integration

📝 Hast Processing

import { createOCRManager } from "unocr";
import tesseractDriver from "unocr/drivers/tesseract";
import aiDriver from "unocr/drivers/ai";
import { toHtml } from "hast-util-to-html";
import { toText } from "hast-util-to-text";
import { rehype } from "rehype";
import { unified } from "unified";

// Works with both Tesseract and AI drivers
const tesseractOcr = createOCRManager({
  driver: tesseractDriver({ langs: "eng" }),
});

const aiOcr = createOCRManager({
  driver: aiDriver({ model: openai(modelName) }),
});

const processor = rehype();

// Process with Tesseract
const tesseractResult = await tesseractOcr.recognize(image);
const tesseractHtml = toHtml(tesseractResult);

// Process with AI
const aiResult = await aiOcr.recognize(image);
const aiHtml = toHtml(aiResult);

// Both outputs are compatible hast format
const text = await processor.process(aiResult);
const processed = unified().use(myPlugin).processSync(tesseractResult);

🔗 Related

• Tesseract.js - JavaScript OCR library
• Vercel AI SDK - AI model integration toolkit
• undio - Universal I/O library
• Hast - HTML Abstract Syntax Tree
• unjs - JavaScript ecosystem

📄 License

MIT © By Sages