An open-source document parsing model with SOTA accuracy and speed, using an RSI loop during post-training.
import os
from typing import Any, Optional
from urllib.parse import urlparse
from pydantic import BaseModel, ConfigDict, Field, field_validator, model_validator
from pydantic_settings import BaseSettings, SettingsConfigDict
import yaml
def _parse_env_bool(value: Optional[str], default: bool = False) -> bool:
if value is None:
return default
return value.strip().lower() in {"true", "1", "yes", "on"}
def _validate_http_url(value: Optional[str]) -> Optional[str]:
if value is None:
return value
parsed = urlparse(value)
if parsed.scheme not in {"http", "https"} or not parsed.netloc:
raise ValueError("must be an absolute http:// or https:// URL")
return value
class ServerConfig(BaseModel):
model_config = ConfigDict(validate_assignment=True)
host: str = "0.0.0.0"
port: int = Field(default=8000, gt=0, le=65535)
workers: int = Field(default=1, gt=0)
class LayoutAnalyzeModelConfig(BaseModel):
model_config = ConfigDict(validate_assignment=True)
url: Optional[str] = None
batch_size: int = Field(default=8, gt=0)
max_concurrency: int = Field(default=2, gt=0)
timeout: int = Field(default=60, gt=0)
retry_times: int = Field(default=2, gt=0)
score_threshold: float = Field(default=0.5, ge=0.0, le=1.0)
nms_iou_threshold: Optional[float] = Field(default=0.6, ge=0.0, le=1.0)
provides_reading_order: bool = False
_validate_url = field_validator("url")(_validate_http_url)
class OXRModelConfig(BaseModel):
model_config = ConfigDict(validate_assignment=True)
url: Optional[str] = None
model_name: str = "oxr"
max_tokens: int = Field(default=8192, gt=0)
max_concurrency: int = Field(default=16, gt=0)
timeout: int = Field(default=120, gt=0)
retry_times: int = Field(default=2, gt=0)
_validate_url = field_validator("url")(_validate_http_url)
class PipelineConfig(BaseModel):
model_config = ConfigDict(validate_assignment=True, extra="forbid")
image_output_dir: str = "outputs"
full_overlap_dedupe: bool = False
table_image_placeholders: bool = False
table_image_min_score: float = Field(default=0.5, ge=0.0, le=1.0)
class Settings(BaseSettings):
debug: bool = Field(default_factory=lambda: _parse_env_bool(os.getenv("DEBUG"), False))
server: ServerConfig = Field(default_factory=ServerConfig)
layout_analyze_model: LayoutAnalyzeModelConfig = Field(default_factory=LayoutAnalyzeModelConfig)
oxr_model: OXRModelConfig = Field(default_factory=OXRModelConfig)
pipeline: PipelineConfig = Field(default_factory=PipelineConfig)
model_config = SettingsConfigDict(
env_prefix="OXR__",
env_nested_delimiter="__",
extra="forbid",
)
@model_validator(mode="before")
@classmethod
def reject_legacy_model_configuration(cls, data: Any) -> Any:
if isinstance(data, dict) and "model" in data:
raise ValueError(
"The legacy 'model' configuration is no longer supported; "
"migrate it to 'layout_analyze_model' and 'oxr_model'."
)
legacy_environment = sorted(
name
for name in os.environ
if name.upper().startswith("OXR_MODEL__") or name.upper().startswith("OXR__MODEL__")
)
if legacy_environment:
raise ValueError(
"legacy model configuration environment variables are not supported "
f"({', '.join(legacy_environment)}); migrate to "
"OXR__LAYOUT_ANALYZE_MODEL__* and OXR__OXR_MODEL__*."
)
return data
def validate_for_server(self) -> None:
missing = []
if not self.layout_analyze_model.url:
missing.append("layout_analyze_model.url")
if not self.oxr_model.url:
missing.append("oxr_model.url")
if missing:
raise ValueError(
"Both model URLs are required to start the server: " + ", ".join(missing)
)
@classmethod
def load_from_yaml(cls, yaml_path: str) -> "Settings":
with open(yaml_path, "r", encoding="utf-8") as f:
data = yaml.safe_load(f) or {}
return cls(**data)
settings = Settings()