Docker Build Configuration#

RAGFlow ships two distinct Docker build strategies: a three-stage Dockerfile for the main application and a single-stage Dockerfile_deepdoc_oss for the standalone DeepDoc microservice.

Main Application: Three-Stage Dockerfile#

The main Dockerfile uses a three-stage pattern:

base stage (lines 1–183):

• Mounts infiniflow/ragflow_deps:latest (a pre-built deps image) to copy model files, NLTK data, Tika JARs, uv, Chrome, and arch-specific binaries without pulling them at build time.
• Installs system packages (libglib2.0-0, default-jdk, ghostscript, pandoc, texlive-*, etc.), nginx, and Node.js 20.x.
• Runs uv python install 3.13 to install the pinned Python interpreter.
• Performs uname -m guards to select arch-correct binaries (uv, stagehand server, MS SQL ODBC driver msodbcsql17 vs msodbcsql18, libssl1.1 deb) for x86_64 vs aarch64.

builder stage (lines 187–243):

• Adds build-essential, libpython3-dev, libicu-dev, libgbm-dev for C extension compilation — kept out of base to shrink the production image.
• Runs uv sync --python 3.13 --frozen against the locked pyproject.toml/uv.lock, with mirror rewriting for China CI (NEED_MIRROR=1). Always re-fetches litellm to avoid stale cached wheels.
• Builds the frontend: npm install then npm run build with VITE_MINIFY=esbuild.
• Writes a VERSION file from git describe.

production stage (lines 246–287):

• Inherits base (no build tools).
• Copies only the compiled .venv, application source (api, rag, deepdoc, agent, mcp, etc.), nginx configs, and the built web/dist from builder.
• Entrypoint: ./entrypoint.sh.

The NEED_MIRROR=1 build arg switches APT, PyPI, and npm to Aliyun/Gitee mirrors throughout all stages 1.

DeepDoc OSS: Single-Stage Dockerfile_deepdoc_oss#

Dockerfile_deepdoc_oss is a minimal ONNX-only image with no torch, no pdfplumber, and no Ascend dependencies:

• Base: ubuntu:24.04 (single stage) 2
• Python: installs python3.12 and python3.12-venv from Ubuntu's apt repos — pinned to 3.12, not 3.13 3
• Dependencies: a small pip install of litserve onnxruntime opencv-python-headless numpy pillow pyclipper python-multipart shapely six huggingface_hub 4
• Stubs: runs docker_stubs.py to generate lightweight stub packages so vision modules load without heavy imports
• Models: downloaded from HuggingFace (or hf-mirror.com with NEED_MIRROR=1) into /app/rag/res/deepdoc 5
• Port: exposes 9390 with a curl-based health check 6
• Entrypoint: python3 deepdoc_server.py --model-dir /app/rag/res/deepdoc 7

In docker-compose.yml, this image is used by the deepdoc profile service, built from the repo root with dockerfile: Dockerfile_deepdoc_oss 8.

Python Version Pinning#

Python version is pinned explicitly across multiple config files:

The DeepDoc server excludes Python 3.13 — its pyproject.toml caps at <3.13 — which is why its Dockerfile installs 3.12 from apt rather than using uv.

Docker Compose Profiles#

docker/docker-compose.yml exposes three profiles via ${RAGFLOW_IMAGE} 13:

Key Files#