ai-form-response-extractor

Hybrid paper + digital form collection powered by multimodal LLMs.

Design a form once, collect responses online and on paper, then use AI to extract structured data from scanned/photographed paper forms and merge everything together.

A lightweight, open-source alternative to enterprise IDP solutions like Rossum, ABBYY FlexiCapture, and Hyperscience.

Features

• SurveyJS-first — First-class adapter for SurveyJS JSON form definitions
• Multi-provider LLMs — OpenAI, Anthropic, Ollama (local models) out of the box
• Intelligent extraction — Text, checkboxes, tables, handwriting from scanned forms
• Multi-page extraction — Pass an ordered array of page images for multi-page paper forms
• Native PDF extraction — Pass digital PDFs directly to providers that support document inputs
• QR / unique ID detection — Automatic form identification from images
• Confidence scoring — Flag low-confidence fields for human review
• Response merging — Combine online + paper responses by unique ID
• Schema-aware prompting — LLM outputs validated against your form schema with Zod

Installation

npm install ai-form-response-extractor

Quick Start

import { createExtractor } from 'ai-form-response-extractor';
import { openai } from 'ai-form-response-extractor/providers';
import { readFileSync } from 'fs';

// 1. Create an extractor with your preferred LLM provider
const extractor = createExtractor({
  provider: openai('gpt-4o'),
  adapter: 'surveyjs',
  options: {
    confidenceThreshold: 0.75,
    maxRetries: 2,
  }
});

// 2. Load your form input (scanned image(s) or native PDF) and form definition
const image = [
  readFileSync('./scanned-form-page-1.png'),
  readFileSync('./scanned-form-page-2.png'),
];
const formDefinition = JSON.parse(readFileSync('./survey.json', 'utf-8'));

// 3. Extract structured data from the provided form input
const result = await extractor.extractFromImage({
  image,
  formDefinition,
});

console.log(result.data);          // Structured responses matching schema
console.log(result.uniqueId);      // Detected QR / barcode ID
console.log(result.confidence);    // Per-field confidence scores

// Single-page forms are also supported:
// image: readFileSync('./scanned-form.png')
// Native PDF is also supported for providers with document input support:
// image: readFileSync('./digital-form.pdf')

PDF Provider Notes

• OpenAI provider: supports native PDF input.
• Anthropic provider: supports native PDF input.
• Ollama provider: current API path is image-only and does not accept native PDF input.

Switching Providers

import { openai, anthropic, ollama } from 'ai-form-response-extractor/providers';

// OpenAI
createExtractor({ provider: openai('gpt-4o') });

// Anthropic
createExtractor({ provider: anthropic('claude-4-sonnet') });

// Local with Ollama (no API key needed)
createExtractor({ provider: ollama('llama-3.2-vision') });

Standalone Utilities

import { detectUniqueId, mergeResponses } from 'ai-form-response-extractor';

// Detect QR code or unique ID from an image
const id = await detectUniqueId(imageBuffer);

// Merge online and paper responses
const merged = mergeResponses(onlineResponses, paperExtractions);

Adapters

| Adapter | Description | |---------|-------------| | surveyjs | Converts SurveyJS JSON into optimized LLM prompts | | json-schema | Standard JSON Schema support | | custom | Bring your own adapter via a simple interface |

Limitations

• signature and signaturepad fields are intentionally not extracted.
• Signatures belong to the source document context and do not make sense as standalone structured data without the original document.

Environment Variables

| Variable | Description | |----------|-------------| | OPENAI_API_KEY | OpenAI API key | | ANTHROPIC_API_KEY | Anthropic API key | | OLLAMA_BASE_URL | Ollama server URL (default: http://localhost:11434) |

Demo

See the ai-form-response-extractor-demo repository for a full working demo.

Documentation

• SPEC.md — Full project specification
• npm
• docs/build-plan.md — Build plan and milestones
• docs/architecture.md — Architecture details
• examples/ — Working examples

Development

# Install dependencies
npm install

# Build
npm run build

# Run tests
npm test

# Lint
npm run lint

Contributing

Contributions are welcome! Please read the spec and build plan before starting work.

1. Fork the repository
2. Create a feature branch (git checkout -b feature/my-feature)
3. Commit your changes (git commit -m 'Add my feature')
4. Push to the branch (git push origin feature/my-feature)
5. Open a Pull Request

License

MIT