Skip to contents

Getting Started with ohseer

unified-interface.Rmd

Introduction

The ohseer package provides a unified interface to multiple OCR (Optical Character Recognition) APIs through a single function: ohseer_ocr(). This means you can switch between different OCR providers without changing your code.

Supported providers:

  • Claude Opus 4.6/Sonnet 4.5: #1 on OCR Arena, structured outputs with JSON schemas
  • Tensorlake: Highest accuracy (91.7%), best for tables and forms
  • Mistral OCR 3: Native markdown output, cost-effective
  • AWS Textract: Reliable option for structured data extraction

Installation 

# Using pak (recommended)
pak::pak("n8layman/ohseer")

# Using devtools
devtools::install_github("n8layman/ohseer")

Authentication

Set up API keys as environment variables:

# Option 1: Set for current session
Sys.setenv(
  ANTHROPIC_API_KEY = "your-claude-key",
  TENSORLAKE_API_KEY = "your-tensorlake-key",
  MISTRAL_API_KEY = "your-mistral-key"
)

Option 2: Create a .env file (recommended):

# .env
ANTHROPIC_API_KEY=your-claude-key
TENSORLAKE_API_KEY=your-tensorlake-key
MISTRAL_API_KEY=your-mistral-key

Getting API Keys

Basic Usage

Process a Document 

library(ohseer)

# Process with default provider (Tensorlake)
result <- ohseer_ocr("document.pdf")

# Access the extracted pages
pages <- result$pages

# Check which provider was used
result$provider

Choose a Provider 

# Use Claude for highest accuracy
result <- ohseer_ocr("document.pdf", provider = "claude")

# Use Mistral for cost-effectiveness
result <- ohseer_ocr("document.pdf", provider = "mistral")

# Use Tensorlake for best table extraction
result <- ohseer_ocr("document.pdf", provider = "tensorlake")

Select Specific Pages 

# Process only first 2 pages
result <- ohseer_ocr("document.pdf", pages = c(1, 2))

# Process specific pages
result <- ohseer_ocr("document.pdf", pages = c(1, 5, 10))

Provider Fallback

Automatically try multiple providers until one succeeds:

# Try providers in order: Tensorlake → Mistral → Claude
result <- ohseer_ocr(
  "document.pdf",
  provider = c("tensorlake", "mistral", "claude")
)

# Check which provider succeeded
message("Used provider: ", result$provider)

# Check error log if any providers failed
if (!is.na(result$error_log)) {
  errors <- jsonlite::fromJSON(result$error_log)
  print(errors)
}

Note: Only providers with API keys set will be tried. Providers without keys are automatically skipped with a warning.

Provider-Specific Options

Each provider supports custom options via ...:

Mistral Options 

result <- ohseer_ocr(
  "document.pdf",
  provider = "mistral",
  extract_header = TRUE,      # Extract headers separately
  extract_footer = TRUE,      # Extract footers separately
  table_format = "markdown"   # Table output format
)

Claude Options 

result <- ohseer_ocr(
  "document.pdf",
  provider = "claude",
  model = "claude-sonnet-4-5",  # Use Sonnet instead of Opus
  max_tokens = 16000,           # Maximum tokens
  schema = my_custom_schema     # Custom JSON schema
)

Tensorlake Options 

result <- ohseer_ocr(
  "document.pdf",
  provider = "tensorlake",
  model = "high-quality-v1",    # Model selection
  timeout = 120                 # Timeout in seconds
)

Understanding Results

Result Structure 

result <- ohseer_ocr("document.pdf")

# Structure:
result$provider   # Character: which provider was used
result$pages      # List: extracted page data
result$raw        # List: raw API response
result$error_log  # Character (JSON): errors from failed providers

Accessing Pages

Different providers return pages in different formats:

# Tensorlake format
result <- ohseer_ocr("doc.pdf", provider = "tensorlake")
page1 <- result$pages[[1]]
page1$page_number      # Integer (1-based)
page1$page_fragments   # List of content fragments

# Mistral format
result <- ohseer_ocr("doc.pdf", provider = "mistral")
page1 <- result$pages[[1]]
page1$index      # Integer (0-based)
page1$markdown   # Full page content as markdown
page1$tables     # Array of table objects

# Claude format (Tensorlake-compatible by default)
result <- ohseer_ocr("doc.pdf", provider = "claude")
page1 <- result$pages[[1]]
page1$page_number      # Integer (1-based)
page1$page_fragments   # List of content fragments

For detailed information on each provider’s output format, see:

Provider Comparison

Provider Accuracy Cost Best For
Claude Opus 4.6 ⭐⭐⭐⭐⭐ High Structured outputs, custom schemas
Tensorlake ⭐⭐⭐⭐⭐ $0.01/page Tables, forms, batch processing
Mistral OCR 3 ⭐⭐⭐ Low Markdown output, cost-sensitive
AWS Textract ⭐⭐⭐⭐ Medium AWS ecosystem

Common Patterns

Strategy 1: Quality First with Fallback 

# Try highest quality first, fall back to cost-effective
result <- ohseer_ocr(
  "document.pdf",
  provider = c("tensorlake", "mistral")
)

Strategy 2: Cost-Optimized with Fallback 

# Try lowest cost first, fall back to higher quality
result <- ohseer_ocr(
  "document.pdf",
  provider = c("mistral", "tensorlake", "claude")
)

Strategy 3: Process Multiple Documents 

files <- c("doc1.pdf", "doc2.pdf", "doc3.pdf")

results <- lapply(files, function(file) {
  tryCatch(
    ohseer_ocr(file, provider = "mistral"),
    error = function(e) {
      warning("Failed to process ", file, ": ", e$message)
      NULL
    }
  )
})

# Filter out failures
successful_results <- Filter(Negate(is.null), results)

Tips

  1. Start simple: Use default provider and settings, then customize as needed

    result <- ohseer_ocr("document.pdf")

  2. Process fewer pages: Reduce costs by processing only the pages you need

    result <- ohseer_ocr("paper.pdf", pages = c(1, 2))

  3. Use provider fallback: Ensure reliability with automatic failover

    result <- ohseer_ocr("doc.pdf", provider = c("tensorlake", "mistral"))

  4. Check provider used: Verify which provider succeeded

    message("Provider: ", result$provider)

  5. Store API keys securely: Use .env file and add to .gitignore

Advanced Topics

For advanced usage, see:

Provider-Specific Functions

While ohseer_ocr() is recommended, provider-specific functions are still available for advanced use:

See the Function Reference for details.