Skip to contents

Process Document with olmOCR

olmocr_ocr.Rd

Runs a document through olmOCR, an open-weights 7B vision-language OCR model from the Allen Institute for AI (AI2). olmOCR converts PDFs and images into clean, natural-reading-order markdown, removing headers and footers and handling equations, tables, handwriting, and multi-column layouts.

Usage

olmocr_ocr(
  file_path,
  api_key = Sys.getenv("OLMOCR_API_KEY"),
  server = Sys.getenv("OLMOCR_SERVER", "https://api.deepinfra.com/v1/openai"),
  model = "allenai/olmOCR-2-7B-1025",
  pages = NULL,
  max_tokens = 8000,
  target_longest_image_dim = 1288,
  extraction_prompt = NULL,
  output_file = NULL,
  timeout = 120
)

Arguments

file_path

Character string. Path to a local PDF, PNG, or JPEG file.

api_key

Character string. Provider API key (Bearer token). Default retrieves from environment variable "OLMOCR_API_KEY". May be empty for an unauthenticated local vLLM server.

server

Character string. Base URL of the OpenAI-compatible endpoint. Default retrieves from environment variable "OLMOCR_SERVER", falling back to DeepInfra ("https://api.deepinfra.com/v1/openai").

model

Character string. Model identifier as registered with the provider. Default is "allenai/olmOCR-2-7B-1025". Note that model names differ between providers (e.g. Cirrascale uses "olmOCR-2-7B-1025").

pages

Integer vector. 1-based page numbers to process. If NULL (default), processes all pages.

max_tokens

Integer. Maximum tokens per page response. Default is 8000.

target_longest_image_dim

Integer. Longest-side pixel target used when rendering PDF pages. Default is 1288.

extraction_prompt

Character string. Custom prompt. If NULL, uses the default olmOCR no-anchoring prompt.

output_file

Character string. Optional path to save the JSON response. Default is NULL (no file output).

timeout

Numeric. Per-request timeout in seconds. Default is 120.

Value

List with structured_output$pages containing one entry per page (page_number, raw text), plus the model and server used. Pass to olmocr_extract_pages for a normalized page list.

Details

Because olmOCR is a GPU model, this function calls an existing OpenAI-compatible endpoint rather than running the model locally. Point it at a hosted provider (DeepInfra, Parasail, Cirrascale) or your own vLLM server via the server argument. See the package README for provider setup.

See also

olmocr_extract_pages, ohseer_ocr

Author

Nathan C. Layman

Examples

if (FALSE) { # \dontrun{
# Using DeepInfra (default endpoint), key in OLMOCR_API_KEY
result <- olmocr_ocr("document.pdf")
pages <- olmocr_extract_pages(result)

# Using Parasail
result <- olmocr_ocr(
  "document.pdf",
  server = "https://api.parasail.io/v1",
  model = "allenai/olmOCR-2-7B-1025"
)

# Using a local vLLM server (no key needed)
result <- olmocr_ocr(
  "document.pdf",
  server = "http://localhost:8000/v1",
  api_key = "",
  model = "olmocr"
)
} # }