How It Works

Document Intelligence is an AI-powered extraction system that reads documents using vision models and returns structured data with confidence scores. Upload a PDF, image, Word document, or ZIP archive, define what you want to extract using plain English, and the system builds a strict JSON extraction prompt for the selected model.

This is inspired by Salesforce MuleSoft IDP, but simpler, faster, and wired into a live model registry covering OpenRouter free models, direct SDK models, and AWS Bedrock.

Here's what happens when you upload a document:

The model dropdown is no longer a fixed hard-coded list. It is generated from live provider state and filtered by health checks:

• OpenRouter: only free, vision-capable models are fetched. Paid OpenRouter models are suppressed because Bedrock is the preferred paid path.
• SDK models: Anthropic, OpenAI, and Gemini direct models are tested individually before they are selectable.
• AWS Bedrock: enabled Bedrock models are shown to signed-in users, with ap-southeast-2 as the configured region.
• Anonymous access: unauthenticated users can only select tested, passing OpenRouter free models. Everything else is greyed out.
• Failure handling: if an OpenRouter model fails during normal extraction, it is removed from rotation and a background refresh is triggered.

Save your extraction configurations as templates to reuse them. Each template stores:

• Field definitions — the prompts you want to extract
• Vision model choice — which model processes these documents
• Shareable link — anyone can load your template with one click
• Optional Salesforce mapping — saved field metadata can be used to generate Salesforce custom objects and fields

Templates are retained for 365 days. Share templates with colleagues so they use the exact same extraction logic, API token, and model policy you defined.

Once you save a template, you get a personal API endpoint to submit documents programmatically. No browser needed — perfect for automation, batch processing, or integrating into your application.

Two modes are available:

• Synchronous — submit a file, wait for results (best for single documents, <10 second turnaround)
• Asynchronous (Polling) — submit a file, get a job ID, poll for results (best for batch processing, concurrent submissions)
• JSON submission — send base64 file content to /api/idp/submit/{token}/json, useful for Salesforce Flow and platforms that cannot send multipart forms.

Anonymous users can test the system using only the currently passing OpenRouter free models. Signed-in users can access additional SDK and Bedrock models when those models are healthy.

Paid usage depends on the selected provider, model, document size, and output length:

• OpenRouter free models — public test path, only if the model passes live image and PDF tests.
• Direct SDK models — useful when provider credits are available, but automatically disabled on quota or billing failures.
• AWS Bedrock — preferred paid path, especially for AU-region/data-residency workflows.
• Bedrock pricing — the admin Bedrock page discovers available multimodal models and refreshes AWS pricing metadata.

Built-in templates and saved samples are available to get you started:

• Identity Document — extracts passport/ID info: document type, expiry date, country, ID number
• Utility Bill — captures account number, due date, amount due, service address
• Medical Report — extracts patient name, date, diagnosis, medications, results
• Website Screenshot — OCR and structured extraction from webpage screenshots
• Saved samples — demo files can be filtered by template and loaded directly into the processor

You can customize any template or start from scratch with your own extraction fields.

What works well:

• Clear, legible documents with good contrast (photos or scans)
• Structured or semi-structured formats (forms, invoices, IDs, certificates)
• Documents in major languages (English, Spanish, French, German, Chinese, etc.)
• Hand-written fields that are clear and legible

What doesn't work well:

• Extremely blurry or rotated images
• Heavily redacted or obscured documents
• Documents with very small or pixelated text
• Images with significant background noise or watermarks

Best practices:

• Use clear, well-lit photos of documents (natural lighting preferred)
• Write extraction prompts as specific questions, not generic labels
• Test with a few sample documents before batch processing
• Use tested OpenRouter free models for public experimentation and demos
• Use Bedrock Nova or Bedrock Claude models for paid, repeatable production-style workflows
• If a model is greyed out, wait for the next health scan or choose another selectable model