PDF Intelligence - AI-Powered PDF Analysis, OCR & RAG Preparation

Extract text, tables and insights from PDFs using AI. Includes OCR for scanned documents and RAG-ready chunking.

Features

• AI Document Analysis: Comprehensive analysis with executive summary, document classification, entity extraction, key topics, and action items
• Text Extraction: Extract clean text from PDFs with optional page markers
• AI-Powered OCR: Use Gemini Vision, GPT-4 Vision, or Claude Vision for scanned PDFs
• Table Detection: Detect and extract tabular data with optional AI enhancement
• Metadata Extraction: Get document information like title, author, dates
• RAG Chunking: Split documents into semantically-aware chunks for vector databases
• Semantic Chunking: AI-powered boundary detection for optimal RAG retrieval
• Large Document Support: Memory-efficient streaming for documents with 100+ pages
• Multiple Output Views: Summary, AI Analysis, Quality Report, Metadata, Content, RAG Chunks
• Dual Operation Modes: One Click (zero config) or BYOK (bring your own keys)
• HTTP REST API: External API access in addition to MCP protocol
• Pay-Per-Event Pricing: Transparent PPE pricing with BYOK discounts

Operation Modes

One Click (Default)

Zero configuration. Platform-managed services with standard PPE billing.

• Just works out of the box
• No API keys needed for basic features
• Standard pricing
• AI features require API keys (see BYOK mode)

BYOK (Bring Your Own Keys)

Use your own API keys for AI features and discounted pricing.

• Up to 50% savings on platform fees (configurable discount)
• Provide your own OpenAI, Anthropic, or Gemini API keys
• Auto-detected when any API key is provided
• Required for AI features (OCR, AI extraction, semantic chunking)

AI Features

AI Document Analysis (Automatic)

Comprehensive AI-powered document analysis that runs automatically when AI keys are configured.

• Supported Providers: Gemini (preferred), OpenAI, Anthropic
• Cost: $0.04 per document
• Output Includes:
  • Executive summary
  • Document type classification
  • Key topics and themes
  • Named entity extraction (people, organizations, dates, locations)
  • Action items and recommendations
  • Key findings and insights
  • Language detection

AI-Powered OCR (enableOcr)

Use Vision APIs to extract text from scanned or image-based PDFs.

• Supported Providers: Gemini Vision (preferred), OpenAI (GPT-4V), Anthropic (Claude Vision)
• Cost: $0.03 per page
• Fallback: Tries providers in order until one succeeds
• Note: Requires PDF pages to contain extractable embedded images

AI Table Extraction (useAiExtraction)

Use AI to intelligently detect and structure tables from document text.

• Supported Providers: Gemini (preferred), OpenAI, Anthropic
• Cost: $0.015 per operation
• Benefits: Better accuracy than rule-based detection, handles complex layouts

Semantic Chunking (semanticChunking)

Use AI to find natural semantic boundaries for document chunking.

• Supported Providers: Gemini (preferred), OpenAI, Anthropic
• Cost: $0.015 per document
• Benefits: Improved RAG retrieval by splitting at conceptual breakpoints

Tools

extract_text

Extract text content from a PDF document.

Input Schema:

Parameter	Type	Required	Default	Description
content	string	No*	-	Base64-encoded PDF content
url	string	No*	-	URL to fetch PDF from
pages	number[]	No	all	Specific pages to extract (1-indexed)
preserveLayout	boolean	No	false	Preserve original text layout and spacing
includePageNumbers	boolean	No	true	Include page number markers in output
enableOcr	boolean	No	false	Use AI Vision OCR for scanned PDFs

*Either content or url must be provided.

Output:

{
  "success": true,
  "overview": {
    "summary": "AI-generated executive summary of the document...",
    "documentType": "Technical Report",
    "pageCount": 5,
    "characterCount": 12500
  },
  "data": {
    "text": "Document content here...",
    "pageCount": 5,
    "extractedPages": [1, 2, 3, 4, 5],
    "characterCount": 12500,
    "aiAnalysis": {
      "executiveSummary": "Comprehensive summary...",
      "documentType": "Technical Report",
      "keyTopics": ["topic1", "topic2"],
      "entities": {
        "people": ["John Doe"],
        "organizations": ["Acme Corp"],
        "dates": ["2024-01-15"]
      },
      "actionItems": ["Review section 3", "Follow up on findings"]
    }
  },
  "intelligence": {
    "qualityScore": 95,
    "confidenceLevel": "high",
    "recommendations": []
  },
  "warnings": []
}

extract_tables

Extract tables from a PDF document.

Input Schema:

Parameter	Type	Required	Default	Description
content	string	No*	-	Base64-encoded PDF content
url	string	No*	-	URL to fetch PDF from
pages	number[]	No	all	Specific pages to process
outputFormat	string	No	"json"	Output format: "json", "csv", or "markdown"
detectHeaders	boolean	No	true	Attempt to detect table headers
useAiExtraction	boolean	No	false	Use AI for intelligent table detection

*Either content or url must be provided.

Output:

{
  "success": true,
  "tables": [
    {
      "page": 1,
      "tableIndex": 0,
      "headers": ["Name", "Value", "Date"],
      "rows": [["Item 1", "100", "2024-01-01"]],
      "rowCount": 1,
      "columnCount": 3
    }
  ],
  "tableCount": 1,
  "warnings": []
}

get_metadata

Extract metadata from a PDF document.

Input Schema:

Parameter	Type	Required	Description
content	string	No*	Base64-encoded PDF content
url	string	No*	URL to fetch PDF from

*Either content or url must be provided.

Output:

{
  "success": true,
  "metadata": {
    "title": "Document Title",
    "author": "Author Name",
    "subject": "Subject",
    "keywords": "keyword1, keyword2",
    "creator": "Creator App",
    "producer": "PDF Producer",
    "creationDate": "2024-01-01T12:00:00",
    "modificationDate": "2024-01-15T09:30:00",
    "pageCount": 10,
    "pdfVersion": "1.7",
    "isEncrypted": false,
    "isLinearized": true
  },
  "warnings": []
}

chunk_for_rag

Split PDF content into chunks optimized for RAG (Retrieval Augmented Generation).

Input Schema:

Parameter	Type	Required	Default	Description
content	string	No*	-	Base64-encoded PDF content
url	string	No*	-	URL to fetch PDF from
chunkSize	number	No	1000	Target chunk size in characters (100-10000)
chunkOverlap	number	No	100	Overlap between chunks (0-500)
splitByPage	boolean	No	false	Never split chunks across page boundaries
includeMetadata	boolean	No	true	Include position metadata in chunks
semanticChunking	boolean	No	false	Use AI for semantic boundary detection

*Either content or url must be provided.

Output:

{
  "success": true,
  "chunks": [
    {
      "id": "chunk_0",
      "text": "Chunk content here...",
      "metadata": {
        "pageNumber": 1,
        "chunkIndex": 0,
        "startChar": 0,
        "endChar": 1000,
        "totalChunks": 5
      }
    }
  ],
  "chunkCount": 5,
  "totalCharacters": 4800,
  "averageChunkSize": 960,
  "warnings": []
}

Output Views

When viewing results in the Apify Console, you can switch between different views:

View	Description
Summary	AI-generated executive summary and key insights
AI Analysis	Full AI analysis with entities, topics, and action items
Quality Report	Quality score, confidence level, and recommendations
Metadata	Document metadata (title, author, dates, page count)
Content	Extracted text content and page information
RAG Chunks	Prepared chunks for vector database ingestion
Full Output	Complete raw output with all data

HTTP REST API

When HTTP API is enabled (default), you can call tools via REST endpoints:

Endpoints

Endpoint	Method	Description
/health	GET	Health check
/info	GET	Actor info and available tools
/api	POST	Execute a tool

Example Request

curl -X POST https://your-username--pdf-processor-mcp.apify.actor/api \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_APIFY_TOKEN" \
  -d '{
    "tool": "extract_text",
    "args": {
      "content": "<base64-pdf>",
      "pages": [1, 2, 3],
      "preserveLayout": false,
      "enableOcr": true
    },
    "includeBilling": true
  }'

Example Response

{
  "success": true,
  "result": {
    "success": true,
    "text": "Document content...",
    "pageCount": 3,
    "extractedPages": [1, 2, 3],
    "characterCount": 5000,
    "warnings": ["OCR completed for 3 pages using AI Vision"]
  },
  "billing": {
    "totalCharged": 0.048,
    "operationMode": "byok",
    "discount": 50,
    "events": [
      { "event": "page-processed", "count": 3, "price": 0.0015 },
      { "event": "ocr-page", "count": 3, "price": 0.0225 }
    ]
  },
  "rateLimit": {
    "remaining": 97,
    "limit": 100,
    "resetIn": 60
  }
}

Pricing

Base Events (One Click / BYOK with 50% discount)

Event	One Click	BYOK (50% off)	Description
page-processed	$0.002	$0.001	Per PDF page processed
document-analyzed	$0.01	$0.005	For metadata extraction
rag-chunking	$0.02	$0.01	For RAG chunking operation
api-call	$0.0002	$0.0001	Per HTTP REST API call
large-data-operation	$0.01	$0.005	For PDFs larger than 1MB

AI Events (Requires API Keys)

Event	One Click	BYOK (50% off)	Description
ocr-page	$0.03	$0.015	Per page OCR with Vision API
ai-table-extraction	$0.015	$0.0075	AI table detection
semantic-chunking	$0.015	$0.0075	AI semantic boundary detection
ai-document-analysis	$0.04	$0.02	Comprehensive AI document analysis

Note: AI events require at least one API key (OpenAI, Anthropic, or Gemini) to be configured. In BYOK mode, you pay for the AI API usage directly to the provider, so the BYOK discount applies only to platform fees.

Pricing Examples

Use Case	Events	Estimated Cost
10-page PDF text extraction	10 pages + analysis	~$0.06
50-page PDF with AI analysis	50 pages + AI analysis	~$0.14
RAG preparation (20 pages)	20 pages + chunking	~$0.06
Scanned PDF OCR (5 pages)	5 OCR pages + analysis	~$0.19

Configuration

Input Parameters

Parameter	Type	Default	Description
operationMode	string	"one-click"	"one-click" or "byok"
googleApiKey	string	-	Google API key for Gemini AI (recommended)
openaiApiKey	string	-	OpenAI API key (enables GPT-4V for OCR)
anthropicApiKey	string	-	Anthropic API key (enables Claude Vision)
preferredAiProvider	string	"auto"	"auto" (Gemini first), "gemini", "openai", or "anthropic"
enableHttpApi	boolean	true	Enable HTTP REST API
apiRateLimit	integer	100	Max requests per minute per client
debug	boolean	false	Enable debug logging

Recommended Setup: For best results, provide a Google API key (Gemini). Gemini 2.5 Flash offers excellent performance at competitive pricing and is the default preferred provider.

Limitations

• Maximum file size: 50MB
• Large document handling: Documents with 100+ pages use memory-efficient streaming mode
• Output truncation: Text output limited to 100k characters; RAG chunks limited to 50 in direct output (full data available via Apify dataset API)
• OCR limitations: OCR works best with PDFs that contain embedded images. For PDFs that require rendering (vector graphics, native text rendered as images), a PDF-to-image conversion step may be needed first.
• Rate limit: 100 requests/minute per client (configurable)
• Memory: Recommended 4GB+ for large documents; Actor supports up to 16GB

Error Codes

Code	Description	Retryable
VALIDATION_ERROR	Invalid input parameters	No
INVALID_PDF	PDF file is corrupted or invalid	No
PROCESSING_ERROR	Error during processing	Yes
RESOURCE_LIMIT	File size or page limit exceeded	No
RATE_LIMIT_EXCEEDED	Too many API requests	Yes

Use Cases

1. Contract Analysis with OCR

Extract text from scanned contracts for AI-powered clause analysis.

{
  "tool": "extract_text",
  "args": {
    "url": "https://example.com/scanned-contract.pdf",
    "preserveLayout": true,
    "enableOcr": true
  }
}

2. Invoice Data Extraction with AI

Extract tables from invoices with AI-powered accuracy.

{
  "tool": "extract_tables",
  "args": {
    "content": "<base64-invoice>",
    "detectHeaders": true,
    "useAiExtraction": true
  }
}

3. Research Paper RAG Processing

Chunk research papers with semantic awareness for better retrieval.

{
  "tool": "chunk_for_rag",
  "args": {
    "content": "<base64-paper>",
    "chunkSize": 500,
    "chunkOverlap": 50,
    "semanticChunking": true
  }
}

4. Document Cataloging

Extract metadata for organizing document libraries.

{
  "tool": "get_metadata",
  "args": {
    "content": "<base64-document>"
  }
}

Local Development

# Install dependencies
npm install

# Build
npm run build

# Run locally in Apify standby mode
npm run dev

Connect to Claude Desktop

Add to ~/.config/claude/claude_desktop_config.json:

{
  "mcpServers": {
    "pdf-processor": {
      "url": "https://your-username--pdf-processor-mcp.apify.actor/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_APIFY_TOKEN"
      }
    }
  }
}

Technical Details

• Runtime: Node.js 22
• Transport: stdio (MCP standard) + HTTP REST API
• PDF Libraries: pdf-parse, pdf-lib
• AI SDKs: Google Generative AI (Gemini 2.5 Flash), OpenAI, Anthropic
• Validation: Zod schemas for all inputs
• Memory Management: 7GB heap with garbage collection optimization

Changelog

v3.0.0

• AI Document Analysis: New comprehensive analysis with executive summary, classification, entities, topics, and action items
• Output Views: Added 7 specialized views in Apify Console (Summary, AI Analysis, Quality Report, Metadata, Content, RAG Chunks, Full Output)
• Large Document Support: Memory-efficient streaming for 100+ page documents
• Output Optimization: Text truncation (100k chars) and chunk limiting (50) to prevent OOM
• Updated Pricing: Adjusted pricing for sustainable 70%+ margins
• Gemini 2.5 Flash: Updated to latest Gemini model as preferred AI provider
• Improved Memory Management: 7GB heap limit with garbage collection optimization

v2.1.0

• Added AI-powered OCR using Vision APIs (GPT-4V, Claude Vision, Gemini Vision)
• Added AI-powered table extraction for better accuracy
• Added semantic chunking with AI boundary detection
• Added preferredAiProvider configuration
• Updated pricing model with AI-specific events
• Improved error messages and warnings

v2.0.0

• Added dual operation modes (One Click and BYOK)
• Added HTTP REST API for external clients
• Added BYOK support with configurable discounts
• Added coherent PPE pricing model
• Added rate limiting for HTTP API

v1.0.0

• Initial release with MCP tools

License

MIT