AI PDF & Document Processing with Agent Skills

Five-Stage Document Processing Workflow

The five-stage workflow transforms raw PDF files in an inbox folder into a fully classified, structured, and semantically searchable knowledge base — automatically.

Stage 1: Upload

Documents are placed in a configured inbox folder that Filesystem MCP monitors. The agent detects new files, reads their filenames and sizes, and queues them for processing. For high-volume environments, the inbox can be an S3 bucket or a shared network drive that Filesystem MCP polls on a schedule.

Stage 2: Parse / OCR

For each document, the agent first attempts extraction with PDF Parser Skill. If the extracted text is empty (indicating a scanned image PDF) or contains garbled characters (indicating a corrupted text layer), the agent automatically switches to OCR Skill. OCR Skill runs Tesseract locally or calls a cloud OCR API, returning clean text with confidence scores for each recognized region.

Stage 3: Extract Data

With clean text in hand, the agent applies an extraction template appropriate to the document type. For invoices: vendor name, invoice number, issue date, due date, line items, subtotal, tax, and total. For contracts: parties, effective date, term, key obligations, payment terms, and termination clauses. For reports: title, date, executive summary, key metrics, and recommendations. The extraction template is defined in plain English and the agent maps the document content to the template fields.

Stage 4: Classify

The agent classifies each document by type (invoice, contract, purchase order, report, correspondence) and assigns metadata tags based on the extracted content: the issuing organization, the relevant department, the date range, and any custom classification criteria you define. Classification decisions are logged for audit purposes.

Stage 5: Store / Index

Notion MCP creates a database entry for each document with all extracted fields as properties and a link to the original file. Embedding Skill chunks the full document text, generates vector embeddings, and stores them in a local vector database for semantic search. From this point, the document is immediately discoverable through both structured database queries and natural language questions.

Step-by-Step Setup

Step 1: Set Up Your Inbox and Output Folders

Create a directory structure for the document pipeline:

mkdir -p ~/documents/inbox
mkdir -p ~/documents/processed
mkdir -p ~/documents/output

Step 2: Configure the MCP Skills

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y", "@modelcontextprotocol/server-filesystem",
        "/Users/you/documents"
      ]
    },
    "pdf-parser": {
      "command": "npx",
      "args": ["-y", "pdf-parser-mcp-server"]
    },
    "ocr": {
      "command": "npx",
      "args": ["-y", "ocr-mcp-server"]
    },
    "notion": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-notion"],
      "env": { "NOTION_API_TOKEN": "$NOTION_API_TOKEN" }
    },
    "embedding": {
      "command": "npx",
      "args": ["-y", "embedding-mcp-server"]
    }
  }
}

Step 3: Test with a Single Document

• "Parse the PDF at ~/documents/inbox/sample-invoice.pdf and extract the vendor name, date, and total"
• "List files in the inbox folder" — verifies Filesystem MCP
• "Create a Notion page titled Test Document with a Today property set to today" — verifies Notion MCP

Use Cases

Invoice Processing Automation

"Process all PDFs in the invoices folder, extract vendor name, invoice number, amount due, and due date from each one, and add a row to the Accounts Payable Notion database for each invoice. Flag any invoice over $10,000 for manual review." The agent processes the batch overnight and the finance team arrives in the morning to a fully populated database with only the exception items requiring attention.

Contract Knowledge Base

"Index all contracts in the legal folder using Embedding Skill so I can search them semantically." Once indexed, you can ask: "Which contracts contain an auto-renewal clause?" or "Find all contracts with a governing law of New York" — and get precise answers with page references from across hundreds of documents.

Report Summarization

"Read all quarterly reports in the reports folder and create a summary page in Notion for each one containing the executive summary, top 3 KPIs, and any risks mentioned." The agent extracts and synthesizes the key content from each report, creating a navigable archive of company intelligence without manual reading or note-taking.

Comparison Table

Frequently Asked Questions