docling separation

application/Dockerfile

@@ -28,8 +28,8 @@ RUN wget -q -O - https://sh.rustup.rs | sh -s -- -y
 # Clean up to reduce container size
 RUN apt-get remove --purge -y wget unzip && apt-get autoremove -y && rm -rf /var/lib/apt/lists/*
 
-# Copy requirements.txt
-COPY requirements.txt .
+# Copy requirements manifests
+COPY requirements.txt requirements-docling.txt ./
 
 # Setup Python virtual environment
 RUN python3.12 -m venv /venv
@@ -42,6 +42,15 @@ RUN pip install --no-cache-dir --upgrade pip && \
     pip install --no-cache-dir tiktoken && \
     pip install --no-cache-dir -r requirements.txt
 
+# Optional docling parser engine (DOC_PARSER_ENGINE=docling, OCR for scanned
+# PDFs/images, read_document's structured output) — OFF by default: it pulls
+# the layout/OCR model stack and adds gigabytes to the image. anydoc (in
+# requirements.txt) is the default parser and needs none of it.
+ARG INSTALL_DOCLING=false
+RUN if [ "$INSTALL_DOCLING" = "true" ]; then \
+        pip install --no-cache-dir -r requirements-docling.txt; \
+    fi
+
 # Final Stage
 FROM ubuntu:24.04 as final
 

application/requirements-docling.txt

@@ -0,0 +1,14 @@
+# Optional docling parser engine.
+#
+# Needed only for: DOC_PARSER_ENGINE=docling, OCR of scanned PDFs and images
+# (DOCLING_OCR_ENABLED / DOCLING_OCR_ATTACHMENTS_ENABLED), .adoc/.vtt/.xml
+# source parsing, and the read_document tool's `structured` output. The
+# default anydoc engine needs none of this.
+#
+# Install ON TOP of requirements.txt (torch/transformers stay pinned there,
+# shared with sentence-transformers):
+#   pip install -r requirements.txt -r requirements-docling.txt
+# Docker: build with --build-arg INSTALL_DOCLING=true
+docling==2.119.0
+rapidocr==3.9.2
+onnxruntime==1.28.0

application/requirements.txt

@@ -12,12 +12,10 @@ cryptography==50.0.0
 dataclasses-json==0.6.7
 daytona==0.205.1
 defusedxml==0.7.1
-docling==2.119.0
-# Default document converter (DOC_PARSER_ENGINE=anydoc); docling above is
-# the optional heavy engine for OCR/structured output.
+# Default document converter (DOC_PARSER_ENGINE=anydoc). The docling engine
+# (OCR, layout/table models, read_document's structured output) is an
+# optional extra: pip install -r requirements-docling.txt
 firecrawl-anydoc==0.2.3
-rapidocr==3.9.2
-onnxruntime==1.28.0
 docx2txt==0.9
 ddgs>=8.0.0
 fast-ebook
@@ -98,10 +96,14 @@ sqlalchemy>=2.0,<3
 starlette>=1.0,<2
 tiktoken==0.13.0
 tokenizers==0.22.2
+# torch stays in core: sentence-transformers (local embeddings) needs it even
+# with the docling extra uninstalled.
 torch==2.11.0
 tqdm==4.67.3
-# Capped <5.9.0 to match docling-core's own darwin pin: 5.9+ breaks docling's
-# PDF layout model on Apple Silicon ("Cannot convert a MPS Tensor to float64").
+# Kept in core for sentence-transformers, so installing the optional docling
+# extra never changes the shared model stack. Capped <5.9.0 to match
+# docling-core's own darwin pin: 5.9+ breaks docling's PDF layout model on
+# Apple Silicon ("Cannot convert a MPS Tensor to float64").
 transformers==5.8.1
 typing-extensions==4.16.0
 typing-inspect==0.9.0