fix: isolate VLM pipeline options to avoid cache collisions

docling_jobkit/convert/manager.py

@@ -31,7 +31,12 @@
     VlmPipelineOptions,
 )
 from docling.datamodel.pipeline_options_vlm_model import ApiVlmOptions, InlineVlmOptions
-from docling.document_converter import DocumentConverter, FormatOption, PdfFormatOption
+from docling.document_converter import (
+    DocumentConverter,
+    FormatOption,
+    ImageFormatOption,
+    PdfFormatOption,
+)
 from docling.models.factories import get_ocr_factory
 from docling.pipeline.vlm_pipeline import VlmPipeline
 from docling_core.types.doc import ImageRefMode
@@ -68,12 +73,26 @@ def _hash_pdf_format_option(pdf_format_option: PdfFormatOption) -> bytes:
         data["pipeline_options"] = pdf_format_option.pipeline_options.model_dump(
             serialize_as_any=True, mode="json"
         )
+        data["pipeline_options_type"] = (
+            f"{pdf_format_option.pipeline_options.__class__.__module__}."
+            f"{pdf_format_option.pipeline_options.__class__.__qualname__}"
+        )
+    else:
+        data["pipeline_options_type"] = None
 
     # Replace `pipeline_cls` with a string representation
-    data["pipeline_cls"] = repr(data["pipeline_cls"])
+    pipeline_cls = pdf_format_option.pipeline_cls
+    data["pipeline_cls"] = (
+        f"{pipeline_cls.__module__}.{pipeline_cls.__qualname__}"
+        if pipeline_cls is not None
+        else "None"
+    )
 
     # Replace `backend` with a string representation
-    data["backend"] = repr(data["backend"])
+    backend = pdf_format_option.backend
+    data["backend"] = (
+        f"{backend.__module__}.{backend.__qualname__}" if backend is not None else "None"
+    )
 
     # Serialize the dictionary to JSON with sorted keys to have consistent hashes
     serialized_data = json.dumps(data, sort_keys=True)
@@ -121,9 +140,19 @@ def _create_converter_cache_from_hash(
         @lru_cache(maxsize=cache_size)
         def _get_converter_from_hash(options_hash: bytes) -> DocumentConverter:
             pdf_format_option = self._options_map[options_hash]
+            image_format_option: FormatOption = pdf_format_option
+            if isinstance(pdf_format_option.pipeline_cls, type) and issubclass(
+                pdf_format_option.pipeline_cls, VlmPipeline
+            ):
+                image_format_option = ImageFormatOption(
+                    pipeline_cls=pdf_format_option.pipeline_cls,
+                    pipeline_options=pdf_format_option.pipeline_options,
+                    backend_options=pdf_format_option.backend_options,
+                )
+
             format_options: dict[InputFormat, FormatOption] = {
                 InputFormat.PDF: pdf_format_option,
-                InputFormat.IMAGE: pdf_format_option,
+                InputFormat.IMAGE: image_format_option,
             }
 
             return DocumentConverter(format_options=format_options)
@@ -282,6 +311,27 @@ def _parse_vlm_pdf_opts(
                 request.vlm_pipeline_model_api.model_dump()
             )
 
+        pipeline_options.do_picture_classification = request.do_picture_classification
+        pipeline_options.do_picture_description = request.do_picture_description
+
+        if request.picture_description_local is not None:
+            pipeline_options.picture_description_options = (
+                PictureDescriptionVlmOptions.model_validate(
+                    request.picture_description_local.model_dump()
+                )
+            )
+
+        if request.picture_description_api is not None:
+            pipeline_options.picture_description_options = (
+                PictureDescriptionApiOptions.model_validate(
+                    request.picture_description_api.model_dump()
+                )
+            )
+
+        pipeline_options.picture_description_options.picture_area_threshold = (
+            request.picture_description_area_threshold
+        )
+
         return pipeline_options
 
     # Computes the PDF pipeline options and returns the PdfFormatOption and its hash

tests/test_options.py

@@ -7,6 +7,7 @@
 from docling.backend.docling_parse_v4_backend import DoclingParseV4DocumentBackend
 from docling.backend.pypdfium2_backend import PyPdfiumDocumentBackend
 from docling.datamodel import vlm_model_specs
+from docling.datamodel.base_models import InputFormat
 from docling.datamodel.pipeline_options import (
     PdfBackend,
     PdfPipelineOptions,
@@ -112,6 +113,12 @@ def test_options_cache_key():
     opts.do_picture_description = True
     pipeline_opts = m.get_pdf_pipeline_opts(opts)
     hash = _hash_pdf_format_option(pipeline_opts)
+    assert hash not in hashes
+    hashes.add(hash)
+
+    opts = ConvertDocumentsOptions(pipeline=ProcessingPipeline.VLM)
+    pipeline_opts = m.get_pdf_pipeline_opts(opts)
+    hash = _hash_pdf_format_option(pipeline_opts)
     # pprint(pipeline_opts.pipeline_options.model_dump(serialize_as_any=True))
     assert hash not in hashes
     hashes.add(hash)
@@ -144,3 +151,13 @@ def test_options_cache_key():
     # pprint(pipeline_opts.pipeline_options.model_dump(serialize_as_any=True))
     assert hash not in hashes
     hashes.add(hash)
+
+
+def test_image_pipeline_uses_vlm_pipeline_when_requested():
+    m = DoclingConverterManager(config=DoclingConverterManagerConfig())
+    opts = ConvertDocumentsOptions(pipeline=ProcessingPipeline.VLM)
+    pipeline_opts = m.get_pdf_pipeline_opts(opts)
+    converter = m.get_converter(pipeline_opts)
+    img_opt = converter.format_to_options[InputFormat.IMAGE]
+    assert img_opt.pipeline_cls == VlmPipeline
+    assert isinstance(img_opt.pipeline_options, VlmPipelineOptions)