DeepDoc Model Distribution#

DeepDoc's ONNX models and native dependency artifacts are pre-packaged into a dedicated ragflow_deps Docker image and downloaded at build time — not fetched lazily at runtime. Two separate download scripts serve two different distribution paths: one for baking artifacts into the full deps image, and one for the lightweight standalone DeepDoc OSS container. In embedded mode, the Recognizer base class will fall back to auto-downloading missing models from HuggingFace Hub at runtime.

Artifacts and Their Sources#

Two Download Scripts#

ragflow_deps/download_deps.py — Full Deps Image#

ragflow_deps/download_deps.py downloads every artifact that the infiniflow/ragflow_deps Docker image bakes in. It must be run before building that image 1:

uv run ragflow_deps/download_deps.py [--china-mirrors]
cd ragflow_deps && docker build -f Dockerfile -t infiniflow/ragflow_deps .

The script anchors its CWD to its own directory so all outputs land under ragflow_deps/, matching the paths expected by ragflow_deps/Dockerfile 2.

It calls huggingface_hub.snapshot_download for two HuggingFace repos 3:

• InfiniFlow/deepdoc
• InfiniFlow/text_concat_xgb_v1.0

After downloading archives, it extracts the three native Go libraries (pdfium, pdf_oxide, office_oxide) into ~/ragflow-native-libs/ for use by build.sh during Go compilation 4. The --china-mirrors flag switches all download URLs to Tsinghua/Huawei Cloud/npmmirror alternatives 5.

deepdoc/server/download_deps.py — DeepDoc OSS Container#

deepdoc/server/download_deps.py is a minimal counterpart used only inside Dockerfile_deepdoc_oss. It downloads only the five ONNX files needed for standalone inference 6:

• layout.onnx, det.onnx, rec.onnx, tsr.onnx, ocr.res

It respects the HF_ENDPOINT environment variable to redirect to https://hf-mirror.com for China builds, and skips files that already exist 7.

The ragflow_deps Image#

ragflow_deps/Dockerfile is built FROM scratch — it contains only the artifacts copied from the ragflow_deps/ directory after download_deps.py runs 8. The main Dockerfile mounts this image at build time via --mount=type=bind,from=infiniflow/ragflow_deps:latest,... to copy models, NLTK data, Tika, Chrome, and arch-specific binaries without any network access during the main image build 9.

DeepDoc OSS Container (Dockerfile_deepdoc_oss)#

Dockerfile_deepdoc_oss is a single-stage Ubuntu 24.04 image that:

1. Installs python3.12, onnxruntime, opencv-python-headless, and minimal dependencies via pip 10
2. Runs deepdoc/server/download_deps.py to pull the five ONNX models from HuggingFace into /app/rag/res/deepdoc 11
3. Runs docker_stubs.py to generate stub packages so vision modules load without torch, pdfplumber, or Ascend imports 12
4. Exposes port 9390 and starts deepdoc_server.py --model-dir /app/rag/res/deepdoc 13

Pass --build-arg NEED_MIRROR=1 to switch both pip and HuggingFace downloads to China-accessible mirrors 14.

China Mirror Support#

Both scripts and both Dockerfiles support NEED_MIRROR=1 / --china-mirrors:

Related Files#