Skip to content

configuration.md

Latest commit

 

History

History
94 lines (74 loc) · 5.36 KB

configuration.md

File metadata and controls

94 lines (74 loc) · 5.36 KB
Raw

Configuration

Settings can come from three places. Precedence, highest first:

  1. CLI flag — --llm-model google/gemini-2.5-flash
  2. Environment variable — LLM_MODEL=google/gemini-2.5-flash
  3. ~/.pdf2md.json — JSON object with any of the keys below
  4. Built-in default

Environment variables

Loaded from .env in the current working directory (walking up) and then ~/.env.

Variable Purpose
OPENROUTER_API_KEY OpenRouter API key
ANTHROPIC_API_KEY Anthropic API key
OPENAI_API_KEY OpenAI API key
PDF2MD_BACKEND Default backend: vision or marker
LLM_PROVIDER Default provider
LLM_MODEL Default per-page model
CLEANUP_MODEL Model used for the cleanup pass (vision backend)
CLEANUP_RPM Cap sustained cleanup requests per minute

~/.pdf2md.json

Run pdf2md config init to write a starter file. Unknown keys are rejected.

Key Type Default Notes
output_dir path ./output Output directory
backend vision|marker vision
no_llm bool false [marker] Skip vision pass
llm_provider string openrouter anthropic, openai, or openrouter
llm_model string provider-specific Per-page model
confidence_threshold float 0.60 [marker] Reprocess text below this score
image_mode enum data-only [marker] all, data-only, or none
max_image_tokens int 1024 [marker] Max output tokens per image call
max_page_tokens int 2200 [vision] Max output tokens per page call
dpi int 200 [vision] Rasterization DPI
rasterize_workers int 4 [vision] Rasterization thread-pool workers
cleanup bool true [vision] Run document-wide cleanup pass
cleanup_model string falls back to llm_model [vision] Model for the cleanup pass
crop_figures bool true [vision] Crop and inline figure regions
url_guard_enabled bool true [vision] Reject hallucinated URLs
max_concurrency int 16 Max concurrent LLM calls
requests_per_minute int or null null Cap sustained per-page requests per minute
cleanup_concurrency int 16 [vision] Max concurrent cleanup-section calls
cleanup_rpm int or null null [vision] Cap sustained cleanup requests per minute
dry_run bool false [marker] Show what would be sent to the LLM
verbose bool false Log per-page / per-image progress

Default models per provider (when llm_model is unset):

Provider Model
anthropic claude-sonnet-4-20250514
openai gpt-4o
openrouter anthropic/claude-sonnet-4

Escape hatches

  • --fast — vision backend, no cleanup pass. Lower latency and cost.
  • --raw — vision backend, no cleanup, no figure cropping, no URL guard. Use when debugging extraction.

--fast and --raw are mutually exclusive, and neither can combine with --cleanup.

Examples

~/.pdf2md.json for a cheap-and-fast default:

{
  "backend": "vision",
  "llm_provider": "openrouter",
  "llm_model": "google/gemini-2.5-flash",
  "cleanup": true,
  "cleanup_model": "deepseek/deepseek-v3.2-exp",
  "dpi": 200,
  "max_concurrency": 16
}

Override the model just for one run:

pdf2md convert paper.pdf --llm-model anthropic/claude-sonnet-4

Throttle when you're rate-limited:

pdf2md convert paper.pdf --rpm 60 --cleanup-rpm 30