An open-source document parsing model with SOTA accuracy and speed, using an RSI loop during post-training.
server:
# Host address the HTTP server binds to.
host: 0.0.0.0
# Port number the HTTP server listens on.
port: 8000
# Number of server worker processes to start.
workers: 1
layout_analyze_model:
# Base URL of the external D-FINE layout service (it receives POST /inference).
url: http://127.0.0.1:8081
# Pages in each multipart layout request.
batch_size: 8
# Concurrent D-FINE HTTP attempts.
max_concurrency: 2
# Seconds allowed for each D-FINE HTTP attempt.
timeout: 60
# Total whole-batch HTTP attempts, including the initial attempt.
retry_times: 2
# Keep candidates whose confidence score is at least this value.
score_threshold: 0.5
# Set true only when the layout service returns blocks in reading order.
provides_reading_order: false
oxr_model:
# Model identifier used when sending requests to the OXR OpenAI-compatible API.
model_name: oxr
# Base URL of the OXR OpenAI-compatible model service.
url: http://127.0.0.1:8080/v1
# Maximum number of tokens the model can generate per request.
max_tokens: 8192
# Global concurrent recognition and reading-order HTTP attempts.
max_concurrency: 16
# Seconds allowed for each OXR HTTP attempt.
timeout: 120
# Total HTTP attempts per OXR request, including the initial attempt.
retry_times: 2
pipeline:
# Directory used to save extracted image outputs.
image_output_dir: outputs
# Opt in to all-category cleanup: outer boxes absorb contained boxes first;
# any remaining positive-area overlap keeps the higher LA confidence.
# Equal confidence keeps the earlier block; missing confidence skips comparison.
full_overlap_dedupe: false
# Restore confident Picture blocks fully contained in a Table via numbered
# blue/white <tit> tokens. Incomplete IDs fall back to the original table image.
table_image_placeholders: false
table_image_min_score: 0.5