chore: bump docling

application/core/settings.py

@@ -95,6 +95,7 @@ class Settings(BaseSettings):
     # Pages docling's threaded pipeline buffers in flight; the library
     # default (100) drives worker RSS to ~3 GB on a mid-size PDF.
     DOCLING_PIPELINE_QUEUE_MAX_SIZE: int = 2
+    DOCLING_COMPILE_TORCH_MODELS: bool = False
     DOCLING_TABULAR_MAX_BYTES: int = 2_000_000
     DOCLING_MARKUP_MAX_BYTES: int = 8_000_000
     ATTACHMENT_TEXT_MAX_BYTES: int = 5_000_000

application/parser/document_reader.py

@@ -252,6 +252,9 @@ def _docling_structured(path: Path, *, ocr_enabled: bool, include_tables: bool,
     else:
         from docling.document_converter import DocumentConverter
 
+        from application.parser.file.docling_parser import _apply_inference_settings
+
+        _apply_inference_settings()
         converter = DocumentConverter()
         doc = converter.convert(str(path)).document
         markdown = doc.export_to_markdown()

application/parser/file/docling_parser.py

@@ -43,6 +43,29 @@ def _apply_pipeline_caps(pipeline_options) -> None:
             setattr(pipeline_options, name, value)
 
 
+def _apply_inference_settings() -> None:
+    """Set docling's global torch.compile toggle from ``DOCLING_COMPILE_TORCH_MODELS``.
+
+    docling compiles its layout/table/OCR models with ``torch.compile`` by
+    default. For DocsGPT's one-shot, per-file parses that warmup is never
+    amortized, and it hard-fails wherever TorchInductor cannot build: it emits
+    invalid Metal on Apple Silicon, and its C++ path does not quote the ``-L``
+    library path, so any install directory containing a space fails to build.
+
+    Guarded so docling builds without the inference settings are unaffected.
+    """
+    from application.core.settings import settings
+
+    try:
+        from docling.datamodel.settings import settings as docling_settings
+    except ImportError:  # pragma: no cover - docling always present here
+        return
+
+    inference = getattr(docling_settings, "inference", None)
+    if inference is not None and hasattr(inference, "compile_torch_models"):
+        inference.compile_torch_models = settings.DOCLING_COMPILE_TORCH_MODELS
+
+
 def _tabular_content_size(file: Path) -> int:
     """Effective content size of a tabular file, in bytes.
 
@@ -256,6 +279,8 @@ class DoclingParser(BaseParser):
         )
         from docling.datamodel.pipeline_options import PdfPipelineOptions
 
+        _apply_inference_settings()
+
         pipeline_options = PdfPipelineOptions(
             do_ocr=self.ocr_enabled,
             do_table_structure=self.table_structure,

application/requirements.txt

@@ -12,7 +12,7 @@ cryptography==50.0.0
 dataclasses-json==0.6.7
 daytona==0.203.0
 defusedxml==0.7.1
-docling==2.84.0
+docling==2.119.0
 rapidocr==3.9.2
 onnxruntime==1.28.0
 docx2txt==0.9