How to Extract Data from PDFs with n8n + Ollama (Free Workflow)

Published March 24, 2026 · 12 min read

PDF processing is one of the most tedious tasks in business operations. Invoices, contracts, reports, receipts — they all arrive as PDFs and need to be manually read, copied, and entered into spreadsheets or databases. Cloud document AI services like AWS Textract or Google Document AI charge $1.50–5.00 per 1,000 pages and require sending sensitive documents to external servers.

With n8n and Ollama, you can build an automated PDF data extraction pipeline that runs entirely on your own hardware — zero API costs, unlimited documents, and your sensitive data never leaves your network.

Why Local AI for Document Processing?

For businesses processing hundreds of invoices, contracts, or reports monthly, the savings compound quickly — and the compliance benefits are significant.

Prerequisites

• n8n — self-hosted (Docker or npm). Installation guide
• Ollama — running locally with a capable model. We recommend llama3 or mistral for document processing.
• A PDF-to-text tool — pdftotext (from poppler-utils) or similar

# Install Ollama
curl -fsSL https://ollama.ai/install.sh | sh
ollama pull llama3

# Install PDF text extraction (Ubuntu/Debian)
sudo apt install poppler-utils

# Verify both work
ollama run llama3 "Say hello"
echo "test" | pdftotext - -

The Workflow: Step by Step

This workflow has 5 nodes that form a complete document processing pipeline:

# Send a PDF to the workflow
curl -X POST http://localhost:5678/webhook/pdf-extract \
  -F "file=@invoice.pdf" \
  -F "extract_type=invoice"

# The command node runs:
pdftotext -layout /tmp/uploaded.pdf -

# The -layout flag preserves the document's visual structure,
# which helps the AI understand tables and columns

# Alternative for scanned documents:
tesseract /tmp/uploaded.pdf /tmp/output pdf
pdftotext -layout /tmp/output.pdf -

// Ollama API call in the HTTP Request node
{
  "model": "llama3",
  "prompt": "Extract the following data from this invoice text and return ONLY valid JSON, no explanation:\n\n{{$node['Extract Text'].json.stdout}}\n\nExtract:\n- invoice_number\n- date (YYYY-MM-DD)\n- due_date (YYYY-MM-DD)\n- vendor: {name, address, vat_id}\n- line_items: [{description, quantity, unit_price, total}]\n- subtotal\n- tax_amount\n- total_amount\n- currency\n\nReturn ONLY the JSON object:",
  "stream": false,
  "options": {
    "temperature": 0.1,
    "num_predict": 2048
  }
}

// Function node: Parse and validate
const response = $input.first().json.response;

// Extract JSON from the AI response
let jsonStr = response;
const jsonMatch = response.match(/\{[\s\S]*\}/);
if (jsonMatch) {
  jsonStr = jsonMatch[0];
}

try {
  const data = JSON.parse(jsonStr);

  // Validate required fields
  const required = ['invoice_number', 'total_amount', 'date'];
  const missing = required.filter(f => !data[f]);

  if (missing.length > 0) {
    return [{ json: {
      status: 'partial',
      missing_fields: missing,
      data: data
    }}];
  }

  // Clean up amounts (remove currency symbols, convert to numbers)
  if (data.total_amount) {
    data.total_amount = parseFloat(
      String(data.total_amount).replace(/[^0-9.-]/g, '')
    );
  }

  return [{ json: { status: 'success', data: data } }];
} catch (e) {
  return [{ json: {
    status: 'error',
    message: 'Failed to parse AI response as JSON',
    raw_response: response
  }}];
}

Adapting for Different Document Types

The beauty of using AI for extraction is that you just change the prompt. Here are templates for common document types:

Contracts

Extract from this contract:
- parties: [{name, role, address}]
- effective_date
- termination_date
- key_terms: [string]
- payment_terms
- governing_law
- signatures: [{name, title, date}]

Receipts

Extract from this receipt:
- store_name
- date
- items: [{name, quantity, price}]
- subtotal
- tax
- total
- payment_method

Resumes / CVs

Extract from this resume:
- name
- email
- phone
- location
- summary
- experience: [{company, title, dates, description}]
- education: [{institution, degree, dates}]
- skills: [string]

Performance and Accuracy

We tested this workflow against 50 real-world invoices with varying layouts and formats:

Common Issues and Fixes

AI returns text instead of JSON

Add "Return ONLY valid JSON. No explanation, no markdown, no code blocks." to the end of your prompt. Also set temperature: 0.1.

Numbers are wrong (e.g., $1,234 becomes 1234)

Add explicit formatting instructions: "Keep all numbers in their original format including decimals and thousands separators."

Scanned PDFs return empty text

Install tesseract-ocr and add an OCR step before the AI extraction. The workflow template includes an optional OCR branch for this.

Processing is too slow

Use a smaller model (mistral is 30% faster than llama3), reduce num_predict, or process pages in parallel using n8n's Split In Batches node.

Real-World Use Cases

• Accounts payable automation: Email forwards invoices to n8n → extracts data → creates entries in accounting software
• Contract management: Upload contracts → extract key dates and terms → set renewal reminders automatically
• Expense reporting: Photo of receipt → OCR + AI extraction → auto-fill expense report
• HR document processing: Resume screening → extract skills and experience → score against job requirements
• Compliance monitoring: Scan regulatory documents → extract requirements → compare against internal policies

Extending the Workflow

Once you have the basic extraction pipeline working, here are ways to extend it:

1. Email integration: Connect an IMAP node to automatically process PDF attachments from incoming emails
2. Batch processing: Use a folder watcher to process all PDFs dropped into a directory
3. Multi-language support: Ollama handles multiple languages natively — just specify the expected language in the prompt
4. Duplicate detection: Add a database lookup to flag duplicate invoices before processing
5. Approval workflow: Route low-confidence extractions to a human reviewer via Slack or email

Getting Started

1. Install n8n and Ollama on your machine (see prerequisites above)
2. Import the workflow JSON into your n8n instance
3. Update the Ollama URL in the HTTP Request node (default: http://localhost:11434)
4. Send a test PDF to the webhook endpoint
5. Connect the output to your preferred destination (Sheets, database, etc.)