Skip to contents

Unified OCR Interface for Multiple Providers with Automatic Fallback

ohseer_ocr.Rd

Usage

ohseer_ocr(
  file_path,
  provider = c("tensorlake", "mistral", "claude"),
  pages = NULL,
  timeout = 60,
  extract_pages = TRUE,
  ...
)

Arguments

file_path

Character string. Path to the PDF file to process.

provider

Character vector. OCR provider(s) to use. Can be a single provider or multiple providers for automatic fallback. Valid values:

"tensorlake"

Tensorlake OCR API (default) - Highest accuracy (91.7\ "mistral"Mistral OCR 3 - Lower cost, native markdown format "claude"Claude Opus/Sonnet - Structured outputs with JSON schema

When multiple providers are specified, they will be tried sequentially until one succeeds. Default is c("tensorlake").

pagesInteger vector. Specific page numbers to process. If NULL (default), processes all pages. Page numbers are 1-based.

timeoutNumeric. Maximum wait time in seconds for OCR processing. Default is 60 seconds. Not used for Claude provider.

extract_pagesLogical. If TRUE (default), automatically extracts and returns page data using provider-specific extraction functions. If FALSE, returns raw API response.

...Additional provider-specific arguments passed to the underlying OCR function:

Tensorlake

model, use_cache, etc.

Mistral

extract_header, extract_footer, table_format, etc.

Claude

model, max_tokens, dpi, schema, etc.

If extract_pages = TRUE (default), returns a list with:

provider

Character string naming the provider that succeeded

pages

List of extracted page data (structure varies by provider)

raw

Raw API response (for advanced use)

error_log

Character string (JSON) of failed attempts, or NA if first provider succeeded

If extract_pages = FALSE, returns the raw provider API response. A unified wrapper function that provides a consistent interface across different OCR providers (Tensorlake, Mistral, Claude). This function normalizes parameter names and return structures, making it easier to switch between providers or implement provider fallback logic. Provider-Specific Output FormatsEach provider returns pages in a different native format:Tensorlake - Structured fragments:

  • page_number: Integer (1-based)

  • page_fragments: List of content fragments with type, content, reading_order

Mistral - Native markdown format:

  • index: Integer (0-based)

  • markdown: Full page content

  • tables: Array of table objects

  • header, footer: Separate header/footer fields

Claude - Structured output (Tensorlake-compatible by default):

  • page_number: Integer (1-based)

  • page_fragments: List of content fragments

  • Custom schema can be provided via schema argument

Provider FallbackWhen multiple providers are specified, they are tried sequentially until one succeeds:


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

Provider-specific functions:

Nathan C. Layman