lfx-nightly 0.2.0.dev41__py3-none-any.whl → 0.3.0.dev3__py3-none-any.whl

lfx/base/agents/agent.py

@@ -71,8 +71,8 @@ class LCAgentComponent(Component):
     ]
 
     outputs = [
-        Output(display_name="Agent", name="agent", method="build_agent", hidden=True, tool_mode=False),
         Output(display_name="Response", name="response", method="message_response"),
+        Output(display_name="Agent", name="agent", method="build_agent", tool_mode=False),
     ]
 
     # Get shared callbacks for tracing and save them to self.shared_callbacks
@@ -185,8 +185,10 @@ class LCAgentComponent(Component):
         if "input" not in input_dict:
             input_dict = {"input": self.input_value}
 
-        if hasattr(self, "system_prompt") and self.system_prompt and self.system_prompt.strip():
-            input_dict["system_prompt"] = self.system_prompt
+        # Use enhanced prompt if available (set by IBM Granite handler), otherwise use original
+        system_prompt_to_use = getattr(self, "_effective_system_prompt", None) or self.system_prompt
+        if system_prompt_to_use and system_prompt_to_use.strip():
+            input_dict["system_prompt"] = system_prompt_to_use
 
         if hasattr(self, "chat_history") and self.chat_history:
             if isinstance(self.chat_history, Data):
@@ -272,9 +274,11 @@ class LCAgentComponent(Component):
                 on_token_callback,
             )
         except ExceptionWithMessageError as e:
-            if hasattr(e, "agent_message") and hasattr(e.agent_message, "id"):
-                msg_id = e.agent_message.id
-                await delete_message(id_=msg_id)
+            # Only delete message from database if it has an ID (was stored)
+            if hasattr(e, "agent_message"):
+                msg_id = e.agent_message.get_id()
+                if msg_id:
+                    await delete_message(id_=msg_id)
             await self._send_message_event(e.agent_message, category="remove_message")
             logger.error(f"ExceptionWithMessageError: {e}")
             raise

lfx/base/agents/altk_base_agent.py

@@ -378,9 +378,11 @@ class ALTKBaseAgentComponent(AgentComponent):
                 cast("SendMessageFunctionType", self.send_message),
             )
         except ExceptionWithMessageError as e:
-            if hasattr(e, "agent_message") and hasattr(e.agent_message, "id"):
-                msg_id = e.agent_message.id
-                await delete_message(id_=msg_id)
+            # Only delete message from database if it has an ID (was stored)
+            if hasattr(e, "agent_message"):
+                msg_id = e.agent_message.get_id()
+                if msg_id:
+                    await delete_message(id_=msg_id)
             await self._send_message_event(e.agent_message, category="remove_message")
             logger.error(f"ExceptionWithMessageError: {e}")
             raise

lfx/base/agents/altk_tool_wrappers.py

@@ -513,7 +513,7 @@ class PostToolProcessor(ALTKBaseTool):
                 output = None
                 try:
                     output = middleware.process(input_data, AgentPhase.RUNTIME)
-                except (AttributeError, TypeError, ValueError, RuntimeError) as e:
+                except Exception as e:  # noqa: BLE001
                     logger.error(f"Exception in executing CodeGenerationComponent: {e}")
                 if output is not None and hasattr(output, "result"):
                     logger.info(f"Output of CodeGenerationComponent: {output.result}")

lfx/base/agents/events.py

@@ -388,7 +388,7 @@ async def process_agent_events(
     agent_message = await send_message_callback(message=agent_message)
     # Capture the original message id - this must stay consistent throughout if streaming
     # Message may not contain id if the Agent is not connected to a Chat Output (_should_skip_message is True)
-    initial_message_id = agent_message.id if hasattr(agent_message, "id") else None
+    initial_message_id = agent_message.get_id()
     try:
         # Create a mapping of run_ids to tool contents
         tool_blocks_map: dict[str, ToolContent] = {}

lfx/base/agents/utils.py

@@ -224,6 +224,10 @@ def get_chat_output_sender_name(self) -> str | None:
     if not hasattr(self, "graph") or not self.graph:
         return None
 
+    # Check if graph has vertices attribute (PlaceholderGraph doesn't)
+    if not hasattr(self.graph, "vertices"):
+        return None
+
     for vertex in self.graph.vertices:
         # Safely check if vertex has data attribute, correct type, and raw_params
         if (

lfx/base/composio/composio_base.py

@@ -41,6 +41,58 @@ class ComposioBaseComponent(Component):
 
     default_tools_limit: int = 5
 
+    # Reserved attribute names that conflict with Component base class
+    RESERVED_ATTRIBUTES: set[str] = {
+        # Core component attributes
+        "name",
+        "description",
+        "status",
+        "display_name",
+        "icon",
+        "priority",
+        "code",
+        "inputs",
+        "outputs",
+        "selected_output",
+        # Properties and methods
+        "trace_type",
+        "trace_name",
+        "function",
+        "repr_value",
+        "field_config",
+        "field_order",
+        "frozen",
+        "build_parameters",
+        "cache",
+        "tools_metadata",
+        "vertex",
+        # User and session attributes
+        "user_id",  # Already handled separately but included for completeness
+        "session_id",
+        "flow_id",
+        "flow_name",
+        "context",
+        # Common method names
+        "build",
+        "run",
+        "stop",
+        "start",
+        "validate",
+        "get_function",
+        "set_attributes",
+        # Additional common conflicts
+        "id",
+        "type",
+        "value",
+        "metadata",
+        "logs",
+        "results",
+        "artifacts",
+        "parameters",
+        "template",
+        "config",
+    }
+
     _base_inputs = [
         MessageTextInput(
             name="entity_id",
@@ -623,13 +675,9 @@ class ComposioBaseComponent(Component):
                                 attachment_related_found = True
                                 continue  # Skip individual attachment fields
 
-                            # Handle conflicting field names - rename user_id to avoid conflicts with entity_id
-                            if clean_field == "user_id":
-                                clean_field = f"{self.app_name}_user_id"
-
-                            # Handle reserved attribute name conflicts (e.g., 'status', 'name')
+                            # Handle reserved attribute name conflicts
                             # Prefix with app name to prevent clashes with component attributes
-                            if clean_field in {"status", "name"}:
+                            if clean_field in self.RESERVED_ATTRIBUTES:
                                 clean_field = f"{self.app_name}_{clean_field}"
 
                             action_fields.append(clean_field)
@@ -795,28 +843,16 @@ class ComposioBaseComponent(Component):
                     # Don't add individual attachment sub-fields to the schema
                     continue
 
-                # Handle conflicting field names - rename user_id to avoid conflicts with entity_id
-                if clean_field_name == "user_id":
-                    clean_field_name = f"{self.app_name}_user_id"
+                # Handle reserved attribute name conflicts
+                if clean_field_name in self.RESERVED_ATTRIBUTES:
+                    original_name = clean_field_name
+                    clean_field_name = f"{self.app_name}_{clean_field_name}"
                     # Update the field schema description to reflect the name change
                     field_schema_copy = field_schema.copy()
+                    original_description = field_schema.get("description", "")
                     field_schema_copy["description"] = (
-                        f"User ID for {self.app_name.title()}: " + field_schema["description"]
-                    )
-                elif clean_field_name == "status":
-                    clean_field_name = f"{self.app_name}_status"
-                    # Update the field schema description to reflect the name change
-                    field_schema_copy = field_schema.copy()
-                    field_schema_copy["description"] = f"Status for {self.app_name.title()}: " + field_schema.get(
-                        "description", ""
-                    )
-                elif clean_field_name == "name":
-                    clean_field_name = f"{self.app_name}_name"
-                    # Update the field schema description to reflect the name change
-                    field_schema_copy = field_schema.copy()
-                    field_schema_copy["description"] = f"Name for {self.app_name.title()}: " + field_schema.get(
-                        "description", ""
-                    )
+                        f"{original_name.replace('_', ' ').title()} for {self.app_name.title()}: {original_description}"
+                    ).strip()
                 else:
                     # Use the original field schema for all other fields
                     field_schema_copy = field_schema
@@ -842,12 +878,8 @@ class ComposioBaseComponent(Component):
                 cleaned_required = []
                 for field in flat_schema["required"]:
                     base = field.replace("[0]", "")
-                    if base == "user_id":
-                        cleaned_required.append(f"{self.app_name}_user_id")
-                    elif base == "status":
-                        cleaned_required.append(f"{self.app_name}_status")
-                    elif base == "name":
-                        cleaned_required.append(f"{self.app_name}_name")
+                    if base in self.RESERVED_ATTRIBUTES:
+                        cleaned_required.append(f"{self.app_name}_{base}")
                     else:
                         cleaned_required.append(base)
                 flat_schema["required"] = cleaned_required
@@ -943,9 +975,10 @@ class ComposioBaseComponent(Component):
                                 inp.advanced = True
 
                             # Skip entity_id being mapped to user_id parameter
-                            if inp.name == "user_id" and getattr(self, "entity_id", None) == getattr(
-                                inp, "value", None
-                            ):
+                            # Check both original name and renamed version
+                            if inp.name in {"user_id", f"{self.app_name}_user_id"} and getattr(
+                                self, "entity_id", None
+                            ) == getattr(inp, "value", None):
                                 continue
 
                             processed_inputs.append(inp)
@@ -2422,12 +2455,11 @@ class ComposioBaseComponent(Component):
 
                 # Handle renamed fields - map back to original names for API execution
                 final_field_name = field
-                if field.endswith("_user_id") and field.startswith(self.app_name):
-                    final_field_name = "user_id"
-                elif field == f"{self.app_name}_status":
-                    final_field_name = "status"
-                elif field == f"{self.app_name}_name":
-                    final_field_name = "name"
+                # Check if this is a renamed reserved attribute
+                if field.startswith(f"{self.app_name}_"):
+                    potential_original = field[len(self.app_name) + 1 :]  # Remove app_name prefix
+                    if potential_original in self.RESERVED_ATTRIBUTES:
+                        final_field_name = potential_original
 
                 arguments[final_field_name] = value
 
@@ -2538,7 +2570,7 @@ class ComposioBaseComponent(Component):
                 build_config[fname]["value"] = "" if fname not in self._bool_variables else False
         # Hide any other visible, non-protected fields that look like parameters
         protected = {
-            "code",
+            # Component control fields
             "entity_id",
             "api_key",
             "auth_link",
@@ -2570,6 +2602,11 @@ class ComposioBaseComponent(Component):
             "instance_url",
             "tenant_id",
         }
+        # Add all reserved Component attributes to protected set
+        protected.update(self.RESERVED_ATTRIBUTES)
+        # Also add the renamed versions (with app_name prefix) to protected set
+        for attr in self.RESERVED_ATTRIBUTES:
+            protected.add(f"{self.app_name}_{attr}")
         # Add all dynamic auth fields to protected set
         protected.update(self._auth_dynamic_fields)
         # Also protect any auth fields discovered across all instances

lfx/base/data/cloud_storage_utils.py

@@ -0,0 +1,156 @@
+"""Shared utilities for cloud storage operations (AWS S3 and Google Drive).
+
+This module provides common functionality used by both read and write file components
+to avoid code duplication.
+"""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+
+
+def validate_aws_credentials(component: Any) -> None:
+    """Validate that required AWS S3 credentials are present.
+
+    Args:
+        component: Component instance with AWS credential attributes
+
+    Raises:
+        ValueError: If any required credential is missing
+    """
+    if not getattr(component, "aws_access_key_id", None):
+        msg = "AWS Access Key ID is required for S3 storage"
+        raise ValueError(msg)
+    if not getattr(component, "aws_secret_access_key", None):
+        msg = "AWS Secret Key is required for S3 storage"
+        raise ValueError(msg)
+    if not getattr(component, "bucket_name", None):
+        msg = "S3 Bucket Name is required for S3 storage"
+        raise ValueError(msg)
+
+
+def create_s3_client(component: Any):
+    """Create and return a configured boto3 S3 client.
+
+    Args:
+        component: Component instance with AWS credential attributes
+
+    Returns:
+        boto3 S3 client instance
+
+    Raises:
+        ImportError: If boto3 is not installed
+    """
+    try:
+        import boto3
+    except ImportError as e:
+        msg = "boto3 is not installed. Please install it using `uv pip install boto3`."
+        raise ImportError(msg) from e
+
+    client_config = {
+        "aws_access_key_id": component.aws_access_key_id,
+        "aws_secret_access_key": component.aws_secret_access_key,
+    }
+
+    if hasattr(component, "aws_region") and component.aws_region:
+        client_config["region_name"] = component.aws_region
+
+    return boto3.client("s3", **client_config)
+
+
+def parse_google_service_account_key(service_account_key: str) -> dict:
+    """Parse Google service account JSON key with multiple fallback strategies.
+
+    This function handles various common formatting issues when users paste
+    service account keys, including:
+    - Control characters
+    - Extra whitespace
+    - Double-encoded JSON strings
+    - Escaped newlines in private_key field
+
+    Args:
+        service_account_key: Service account JSON key as string
+
+    Returns:
+        dict: Parsed service account credentials
+
+    Raises:
+        ValueError: If all parsing strategies fail
+    """
+    credentials_dict = None
+    parse_errors = []
+
+    # Strategy 1: Parse as-is with strict=False to allow control characters
+    try:
+        credentials_dict = json.loads(service_account_key, strict=False)
+    except json.JSONDecodeError as e:
+        parse_errors.append(f"Standard parse: {e!s}")
+
+    # Strategy 2: Strip whitespace and try again
+    if credentials_dict is None:
+        try:
+            cleaned_key = service_account_key.strip()
+            credentials_dict = json.loads(cleaned_key, strict=False)
+        except json.JSONDecodeError as e:
+            parse_errors.append(f"Stripped parse: {e!s}")
+
+    # Strategy 3: Check if it's double-encoded (JSON string of a JSON string)
+    if credentials_dict is None:
+        try:
+            decoded_once = json.loads(service_account_key, strict=False)
+            credentials_dict = json.loads(decoded_once, strict=False) if isinstance(decoded_once, str) else decoded_once
+        except json.JSONDecodeError as e:
+            parse_errors.append(f"Double-encoded parse: {e!s}")
+
+    # Strategy 4: Try to fix common issues with newlines in the private_key field
+    if credentials_dict is None:
+        try:
+            # Replace literal \n with actual newlines which is common in pasted JSON
+            fixed_key = service_account_key.replace("\\n", "\n")
+            credentials_dict = json.loads(fixed_key, strict=False)
+        except json.JSONDecodeError as e:
+            parse_errors.append(f"Newline-fixed parse: {e!s}")
+
+    if credentials_dict is None:
+        error_details = "; ".join(parse_errors)
+        msg = (
+            f"Unable to parse service account key JSON. Tried multiple strategies: {error_details}. "
+            "Please ensure you've copied the entire JSON content from your service account key file. "
+            "The JSON should start with '{' and contain fields like 'type', 'project_id', 'private_key', etc."
+        )
+        raise ValueError(msg)
+
+    return credentials_dict
+
+
+def create_google_drive_service(service_account_key: str, scopes: list[str], *, return_credentials: bool = False):
+    """Create and return a configured Google Drive API service.
+
+    Args:
+        service_account_key: Service account JSON key as string
+        scopes: List of Google API scopes to request
+        return_credentials: If True, return both service and credentials as tuple
+
+    Returns:
+        Google Drive API service instance, or tuple of (service, credentials) if return_credentials=True
+
+    Raises:
+        ImportError: If Google API client libraries are not installed
+        ValueError: If credentials cannot be parsed
+    """
+    try:
+        from google.oauth2 import service_account
+        from googleapiclient.discovery import build
+    except ImportError as e:
+        msg = "Google API client libraries are not installed. Please install them."
+        raise ImportError(msg) from e
+
+    credentials_dict = parse_google_service_account_key(service_account_key)
+
+    credentials = service_account.Credentials.from_service_account_info(credentials_dict, scopes=scopes)
+    service = build("drive", "v3", credentials=credentials)
+
+    if return_credentials:
+        return service, credentials
+    return service

lfx/base/data/docling_utils.py

@@ -3,7 +3,7 @@ import signal
 import sys
 import traceback
 from contextlib import suppress
-from typing import TYPE_CHECKING
+from functools import lru_cache
 
 from docling_core.types.doc import DoclingDocument
 from pydantic import BaseModel, SecretStr, TypeAdapter
@@ -12,9 +12,6 @@ from lfx.log.logger import logger
 from lfx.schema.data import Data
 from lfx.schema.dataframe import DataFrame
 
-if TYPE_CHECKING:
-    from langchain_core.language_models.chat_models import BaseChatModel
-
 
 class DoclingDependencyError(Exception):
     """Custom exception for missing Docling dependencies."""
@@ -152,6 +149,81 @@ def _deserialize_pydantic_model(data: dict):
     return adapter.validate_python(data["config"])
 
 
+# Global cache for DocumentConverter instances
+# This cache persists across multiple runs and thread invocations
+@lru_cache(maxsize=4)
+def _get_cached_converter(
+    pipeline: str,
+    ocr_engine: str,
+    *,
+    do_picture_classification: bool,
+    pic_desc_config_hash: str | None,
+):
+    """Create and cache a DocumentConverter instance based on configuration.
+
+    This function uses LRU caching to maintain DocumentConverter instances in memory,
+    eliminating the 15-20 minute model loading time on subsequent runs.
+
+    Args:
+        pipeline: The pipeline type ("standard" or "vlm")
+        ocr_engine: The OCR engine to use
+        do_picture_classification: Whether to enable picture classification
+        pic_desc_config_hash: Hash of the picture description config (for cache key)
+
+    Returns:
+        A cached or newly created DocumentConverter instance
+    """
+    from docling.datamodel.base_models import InputFormat
+    from docling.datamodel.pipeline_options import OcrOptions, PdfPipelineOptions, VlmPipelineOptions
+    from docling.document_converter import DocumentConverter, FormatOption, PdfFormatOption
+    from docling.models.factories import get_ocr_factory
+    from docling.pipeline.vlm_pipeline import VlmPipeline
+
+    logger.info(f"Creating DocumentConverter for pipeline={pipeline}, ocr_engine={ocr_engine}")
+
+    # Configure the standard PDF pipeline
+    def _get_standard_opts() -> PdfPipelineOptions:
+        pipeline_options = PdfPipelineOptions()
+        pipeline_options.do_ocr = ocr_engine not in {"", "None"}
+        if pipeline_options.do_ocr:
+            ocr_factory = get_ocr_factory(
+                allow_external_plugins=False,
+            )
+            ocr_options: OcrOptions = ocr_factory.create_options(
+                kind=ocr_engine,
+            )
+            pipeline_options.ocr_options = ocr_options
+
+        pipeline_options.do_picture_classification = do_picture_classification
+
+        # Note: pic_desc_config_hash is for cache key only
+        # Actual picture description is handled separately (non-cached path)
+        _ = pic_desc_config_hash  # Mark as intentionally unused
+
+        return pipeline_options
+
+    # Configure the VLM pipeline
+    def _get_vlm_opts() -> VlmPipelineOptions:
+        return VlmPipelineOptions()
+
+    if pipeline == "standard":
+        pdf_format_option = PdfFormatOption(
+            pipeline_options=_get_standard_opts(),
+        )
+    elif pipeline == "vlm":
+        pdf_format_option = PdfFormatOption(pipeline_cls=VlmPipeline, pipeline_options=_get_vlm_opts())
+    else:
+        msg = f"Unknown pipeline: {pipeline!r}"
+        raise ValueError(msg)
+
+    format_options: dict[InputFormat, FormatOption] = {
+        InputFormat.PDF: pdf_format_option,
+        InputFormat.IMAGE: pdf_format_option,
+    }
+
+    return DocumentConverter(format_options=format_options)
+
+
 def docling_worker(
     *,
     file_paths: list[str],
@@ -162,7 +234,12 @@ def docling_worker(
     pic_desc_config: dict | None,
     pic_desc_prompt: str,
 ):
-    """Worker function for processing files with Docling in a separate process."""
+    """Worker function for processing files with Docling using threading.
+
+    This function now uses a globally cached DocumentConverter instance,
+    significantly reducing processing time on subsequent runs from 15-20 minutes
+    to just seconds.
+    """
     # Signal handling for graceful shutdown
     shutdown_requested = False
 
@@ -205,12 +282,12 @@ def docling_worker(
     check_shutdown()
 
     try:
-        from docling.datamodel.base_models import ConversionStatus, InputFormat
-        from docling.datamodel.pipeline_options import OcrOptions, PdfPipelineOptions, VlmPipelineOptions
-        from docling.document_converter import DocumentConverter, FormatOption, PdfFormatOption
-        from docling.models.factories import get_ocr_factory
-        from docling.pipeline.vlm_pipeline import VlmPipeline
-        from langchain_docling.picture_description import PictureDescriptionLangChainOptions
+        from docling.datamodel.base_models import ConversionStatus, InputFormat  # noqa: F401
+        from docling.datamodel.pipeline_options import OcrOptions, PdfPipelineOptions, VlmPipelineOptions  # noqa: F401
+        from docling.document_converter import DocumentConverter, FormatOption, PdfFormatOption  # noqa: F401
+        from docling.models.factories import get_ocr_factory  # noqa: F401
+        from docling.pipeline.vlm_pipeline import VlmPipeline  # noqa: F401
+        from langchain_docling.picture_description import PictureDescriptionLangChainOptions  # noqa: F401
 
         # Check for shutdown after imports
         check_shutdown()
@@ -233,27 +310,34 @@ def docling_worker(
         queue.put({"error": "Worker interrupted during imports", "shutdown": True})
         return
 
-    # Configure the standard PDF pipeline
-    def _get_standard_opts() -> PdfPipelineOptions:
+    # Use cached converter instead of creating new one each time
+    # This is the key optimization that eliminates 15-20 minute model load times
+    def _get_converter() -> DocumentConverter:
         check_shutdown()  # Check before heavy operations
 
-        pipeline_options = PdfPipelineOptions()
-        pipeline_options.do_ocr = ocr_engine not in {"", "None"}
-        if pipeline_options.do_ocr:
-            ocr_factory = get_ocr_factory(
-                allow_external_plugins=False,
-            )
-
-            ocr_options: OcrOptions = ocr_factory.create_options(
-                kind=ocr_engine,
-            )
-            pipeline_options.ocr_options = ocr_options
-
-        pipeline_options.do_picture_classification = do_picture_classification
-
+        # For now, we don't support pic_desc_config caching due to serialization complexity
+        # This is a known limitation that can be addressed in a future enhancement
         if pic_desc_config:
-            pic_desc_llm: BaseChatModel = _deserialize_pydantic_model(pic_desc_config)
-
+            logger.warning(
+                "Picture description with LLM is not yet supported with cached converters. "
+                "Using non-cached converter for this request."
+            )
+            # Fall back to creating a new converter (old behavior)
+            from docling.datamodel.base_models import InputFormat
+            from docling.datamodel.pipeline_options import PdfPipelineOptions
+            from docling.document_converter import DocumentConverter, FormatOption, PdfFormatOption
+            from docling.models.factories import get_ocr_factory
+            from langchain_docling.picture_description import PictureDescriptionLangChainOptions
+
+            pipeline_options = PdfPipelineOptions()
+            pipeline_options.do_ocr = ocr_engine not in {"", "None"}
+            if pipeline_options.do_ocr:
+                ocr_factory = get_ocr_factory(allow_external_plugins=False)
+                ocr_options = ocr_factory.create_options(kind=ocr_engine)
+                pipeline_options.ocr_options = ocr_options
+
+            pipeline_options.do_picture_classification = do_picture_classification
+            pic_desc_llm = _deserialize_pydantic_model(pic_desc_config)
             logger.info("Docling enabling the picture description stage.")
             pipeline_options.do_picture_description = True
             pipeline_options.allow_external_plugins = True
@@ -261,33 +345,24 @@ def docling_worker(
                 llm=pic_desc_llm,
                 prompt=pic_desc_prompt,
             )
-        return pipeline_options
 
-    # Configure the VLM pipeline
-    def _get_vlm_opts() -> VlmPipelineOptions:
-        check_shutdown()  # Check before heavy operations
-        return VlmPipelineOptions()
-
-    # Configure the main format options and create the DocumentConverter()
-    def _get_converter() -> DocumentConverter:
-        check_shutdown()  # Check before heavy operations
-
-        if pipeline == "standard":
-            pdf_format_option = PdfFormatOption(
-                pipeline_options=_get_standard_opts(),
-            )
-        elif pipeline == "vlm":
-            pdf_format_option = PdfFormatOption(pipeline_cls=VlmPipeline, pipeline_options=_get_vlm_opts())
-        else:
-            msg = f"Unknown pipeline: {pipeline!r}"
-            raise ValueError(msg)
-
-        format_options: dict[InputFormat, FormatOption] = {
-            InputFormat.PDF: pdf_format_option,
-            InputFormat.IMAGE: pdf_format_option,
-        }
-
-        return DocumentConverter(format_options=format_options)
+            pdf_format_option = PdfFormatOption(pipeline_options=pipeline_options)
+            format_options: dict[InputFormat, FormatOption] = {
+                InputFormat.PDF: pdf_format_option,
+                InputFormat.IMAGE: pdf_format_option,
+            }
+            return DocumentConverter(format_options=format_options)
+
+        # Use cached converter - this is where the magic happens!
+        # First run: creates and caches converter (15-20 min)
+        # Subsequent runs: reuses cached converter (seconds)
+        pic_desc_config_hash = None  # Will be None since we checked above
+        return _get_cached_converter(
+            pipeline=pipeline,
+            ocr_engine=ocr_engine,
+            do_picture_classification=do_picture_classification,
+            pic_desc_config_hash=pic_desc_config_hash,
+        )
 
     try:
         # Check for shutdown before creating converter (can be slow)