This quarter DocumentPro shipped five major upgrades to its AI document extraction platform: semantic chunking that automatically splits multi-invoice PDFs into individual documents, per-field confidence scores on extraction results, a document classifier API, two new frontier AI models (Claude Sonnet 5 and GPT 5.6 Terra) with automatic model fallback, and an MCP server that lets AI agents build on DocumentPro directly. We also shipped signed webhooks, a US tax template library, and a free plan that now renews with 200 free credits every month.

If you process invoices, receipts, purchase orders, or tax documents at any volume — or you're a developer wiring document extraction into a product — here's what each change does, why it exists, and how to try it.

1. Semantic Chunking: Multi-Invoice PDFs Now Split Themselves

Semantic chunking automatically detects document boundaries inside a combined PDF and splits each document into its own record before extraction — no manual pre-splitting, no fixed page-count rules.

Anyone who handles supplier paperwork knows the file: one PDF, a dozen invoices inside, and no pattern to where one ends and the next begins. Invoice #1 runs three pages, invoice #2 runs five, invoice #3 has a remittance slip stapled to it. Until now the answer was manual splitting, or brittle "every N pages" rules that break the first time a vendor adds a page.

DocumentPro now reads the batch file the way a person would. The AI looks at the content of each page — headers, invoice numbers, totals, layout resets — and determines where each document starts and ends. A 12-page vendor_batch.pdf comes back as three separate invoice records, each extracted into its own structured JSON, each with its own status in your dashboard and API.

This matters most for the workflows where batching is the norm:

• Accounts payable teams receiving consolidated vendor statements or scanned mail batches — see how this fits an invoice extraction workflow.
• Tax and accounting firms handling combined client uploads — a single PDF containing a stack of Form 1099s or W-2s splits into one record per form.
• Developers whose users upload whatever they have, in whatever shape it arrives.

There's nothing to configure. Upload a multi-document file to a parser and the splitting happens before extraction. Try it on a real batch file — the free plan's monthly credits cover it.

2. Per-Field Confidence Scores: Know How Sure the Model Is

Every field in an extraction response can now carry its own confidence score — a self-reported 0–1 rating of how certain the model is about that specific value.

Document extraction accuracy is never uniform. On the same invoice, the model might be near-certain about the invoice number and total but shaky on a smudged PO number. Document-level review can't see that difference; you end up either trusting everything or re-checking everything.

Per-field confidence makes the uncertainty visible exactly where it lives:

{
  "invoice_no": { "value": "INV-2041", "confidence": 0.98 },
  "total": { "value": "$12,480.00", "confidence": 0.97 },
  "po_number": { "value": "PO-88712", "confidence": 0.62 }
}

That 0.62 is the whole point. Instead of a human re-reading the full document, your pipeline auto-accepts the two high-confidence fields and flags one field for a quick look. Teams use the same scores to drive model escalation — run an economical model first and automatically re-run only low-confidence documents through a frontier model, a pattern we covered in depth in our guide to choosing an AI model for document extraction.

Per-field confidence scores are enabled on request. Talk to us and we'll switch them on for your account.

3. Classifier API: Sort Any Document in One API Call

The new classifier API identifies what type of document you're looking at before you extract it. Define your categories, send a document, and get back the best match with a confidence score for every label:

{
  "classification": "invoice",
  "confidence_scores": {
    "invoice": 0.94,
    "purchase_order": 0.04,
    "receipt": 0.02
  }
}

Most real document pipelines don't receive one document type — they receive an inbox. Invoices, receipts, credit notes, purchase orders, and contracts arrive through the same upload form or email address, and something has to route each one to the right extraction template. Historically that "something" was a rules engine (fragile) or a person (expensive).

With the classifier API you define your categories once, save the classifier, and route programmatically: classification says invoice, so the file goes to your invoice parser; receipt goes to expenses; anything below your confidence threshold goes to a human. Combined with semantic chunking above, a mixed batch PDF can be split and sorted before a single field is extracted.

The classifier API documentation has the full request format, or see the broader document AI API developer guide for how classification fits into an end-to-end pipeline.

4. New Models: Claude Sonnet 5 and GPT 5.6 Terra, Plus Automatic Fallback

The newest frontier models from Anthropic and OpenAI are now live for extraction in DocumentPro. Claude Sonnet 5 brings substantially higher-fidelity vision on dense, small-print, and degraded pages — exactly what scanned tax packets and photographed receipts need — and GPT 5.6 Terra is the strongest GPT-family option for hard documents. They join GPT 4o, GPT 4o Mini, GPT o4 Mini, and Claude Sonnet 4.6 in the model picker.

Because the model is a parser setting, trying either one takes a single settings change — same fields, same validation rules, same API integration, no prompt engineering. Duplicate a parser, set a different model on the copy, and run the same 20 documents through both to see which earns its credits on your paperwork. Our model selection guide walks through that benchmark step by step, including when a cheaper model is genuinely the right answer.

Extraction also got more resilient this quarter: if a model fails or is temporarily unavailable, DocumentPro now automatically falls back through a chain of alternative models — including Claude Haiku 4.5 — instead of failing your job. Your pipeline keeps moving even when an upstream model provider doesn't.

Test the new models on a parser today.

5. MCP Server: Your AI Agents Can Now Build on DocumentPro

DocumentPro now runs a hosted MCP (Model Context Protocol) server, so AI coding agents can use DocumentPro directly — with exact field schemas instead of guessed endpoint shapes. Point Claude Code, Claude Desktop, Cursor, Replit, Lovable, or any MCP-capable client at the server and it gets 11 tools spanning extraction, classification, and template authoring.

Connecting takes one config block and the same API key you already have:

"documentpro": {
  "url": "https://api.documentpro.ai/mcp",
  "headers": { "x-api-key": "<your-key>" }
}

Why this matters: a growing share of integration code is written by agents, not people. When your agent scaffolds an invoice-processing feature, it can now call DocumentPro's real tools while it works — list your templates, check the exact schema of a parser, run a test extraction — and generate integration code that matches reality on the first pass. If you've been asking an AI assistant to "add document extraction" to your app and watching it hallucinate an API, this is the fix.

Setup instructions for each client are in the MCP server documentation. For the bigger picture on embedding extraction into your product, see how to add document extraction to your SaaS.

6. Also Shipped: Signed Webhooks and DocumentPro for Tax

Two more upgrades from this quarter that deserve a mention:

• Signed webhooks. Every webhook payload can now be verified with an HMAC-SHA256 signature, so your endpoint can cryptographically confirm that a "document processed" event really came from DocumentPro — a must-have when webhook events trigger payments or postings downstream.
• DocumentPro for Tax. A library of 35 pre-built templates for US tax documents — W-2s, 1099s (including the multi-variant 1099 family), K-1s, and more — so tax and accounting teams can start extracting on day one instead of building schemas form by form. Combined with semantic chunking, a combined client upload of mixed tax forms splits and extracts automatically. See the tax and accounting solution for the full workflow.

7. The Free Plan Now Renews: 200 Credits Every Month

Every DocumentPro account now includes 200 free credits every month — recurring, not a one-time trial bundle. If you signed up a while ago, the renewing credits are already active in your account.

This changes what the free plan is for. A one-time bundle lets you evaluate once; a monthly allowance lets you run something — a low-volume extraction workflow, a monthly model benchmark, an agent hooked up to the MCP server — indefinitely, on every model in the lineup including Claude Sonnet 5 and GPT 5.6 Terra. When your volume outgrows 200 credits, pricing scales from there.

Claim your monthly credits — signup takes a minute and doesn't require a card.

Frequently Asked Questions

How do I split a PDF that contains multiple invoices? You no longer need to split it yourself. DocumentPro's semantic chunking detects document boundaries inside a combined PDF with AI — even when each invoice has a different page count — and splits each one into its own record before extraction. Upload the batch file as-is and every invoice comes back as a separate, fully extracted document.

What are confidence scores in document extraction? A confidence score is a 0–1 rating of how certain the AI model is about an extracted value. DocumentPro can attach a score to every individual field in the JSON response, so instead of manually spot-checking whole documents you route only low-confidence fields — say, anything under 0.8 — to human review. Per-field confidence is enabled on request for your account.

How do I classify documents automatically with an API? Define your categories once (for example invoice, receipt, purchase order), save them as a classifier in DocumentPro, and send each incoming document in a single API call. The response returns the best-matching category plus a confidence score for every label, which lets you route documents to the right extraction workflow automatically — no rules engine to build or maintain.

Which AI models does DocumentPro support? DocumentPro currently offers six models for extraction, including the newly added Claude Sonnet 5 and GPT 5.6 Terra alongside GPT 4o, GPT 4o Mini, GPT o4 Mini, and Claude Sonnet 4.6. The model is a parser setting, so you can switch or A/B test models without changing your fields, API calls, or webhooks. If a model fails, extraction automatically falls back through a chain of alternative models instead of failing your job.

What is the DocumentPro MCP server? It is a hosted Model Context Protocol server at api.documentpro.ai/mcp that gives AI agents and coding assistants — Claude Code, Claude Desktop, Cursor, Replit, Lovable, and any other MCP-capable client — direct access to 11 DocumentPro tools covering extraction, classification, and template authoring, with exact field schemas. It authenticates with the same API key you already use.

Does DocumentPro have a free plan? Yes, and it got significantly better this quarter: every account now receives 200 free credits every month, recurring — not a one-time trial bundle. That is enough to run real extraction, classification, and model benchmarks each month before deciding to upgrade, and existing accounts get the renewing credits automatically.

Conclusion

The thread running through this quarter's release is less manual work around the extraction itself: batch files split themselves, uncertain fields flag themselves, documents sort themselves to the right workflow, failed models replace themselves, and your AI agents integrate DocumentPro without you writing the glue code. And with 200 credits renewing every month, all of it is testable on the free plan before you spend anything.

Sign up and see what's new in your account, or if you'd rather walk through your document workflow with us — batch splitting for accounts payable, the tax template library, or enabling per-field confidence scores — talk to us.

Also Read: How to Choose the Right AI Model for Document Extraction | Document AI API: The Developer Guide | Invoice Data Extraction: How to Extract Data from Invoices