Dockerfile.standalone

# ModelSEED API: self-contained image for standalone use
#
# Unlike the production Dockerfile (which COPYs sibling repos from the
# build host), this image clones all dependency repos from GitHub during
# the build. Result: a single self-contained image that anyone can pull
# from ghcr.io and run without setting up a local development environment.
#
# Build (CI): GitHub Actions builds + pushes to ghcr.io/modelseed/modelseed-api
# Build (local, for testing):
#     docker build -f Dockerfile.standalone -t modelseed-api:standalone .
# Run:
#     docker run -p 8000:8000 ghcr.io/modelseed/modelseed-api:latest
#     # then hit http://localhost:8000/demo/
#
# The image includes ModelSEEDDatabase and ModelSEEDTemplates baked in;
# total size ~1.5-2 GB. Larger than typical web-app images but eliminates
# the need for separate data downloads.
#
# Defaults to local-storage mode (no PATRIC account needed). Workspace routes
# use that local backend; hosted-only RAST integrations return 503 unless their
# environment variables are configured. See docs/STANDALONE.md.

FROM python:3.11-slim

# System deps: GLPK for cobra's linear solver, git for cloning, gcc/g++
# for Python wheel builds that need compilation.
RUN apt-get update && apt-get install -y --no-install-recommends \
        glpk-utils \
        libglpk-dev \
        libexpat1 \
        gcc \
        g++ \
        git \
    && rm -rf /var/lib/apt/lists/*

WORKDIR /deps

# Clone dependency repos, pinned to specific commits (not just branch
# names) so a rebuild of this same Dockerfile later reproduces the exact
# code archived under this release's DOI, even after upstream branches
# move on. Each uses a shallow fetch-by-SHA (supported by GitHub's
# uploadpack.allowReachableSHA1InWant) to keep image size down without
# giving up pinning. Update these SHAs deliberately when bumping deps;
# `git ls-remote <repo-url> refs/heads/<branch>` gets the latest.
RUN pin_clone() { \
        mkdir "$2" && cd "$2" && git init -q && \
        git remote add origin "$1" && \
        git fetch -q --depth 1 origin "$3" && \
        git checkout -q FETCH_HEAD && cd ..; \
    }; \
    pin_clone https://github.com/Fxe/cobrakbase.git cobrakbase e7bbd3c37b96d44ac49b7b16dfa6fc08d355d4a7 && \
    pin_clone https://github.com/cshenry/ModelSEEDpy.git ModelSEEDpy b758b44217bca756a5e07a70e986fbdfe73dd1b6 && \
    pin_clone https://github.com/cshenry/KBUtilLib.git KBUtilLib 0f8480327b52b208ea8aa366b9c738576162d912 && \
    pin_clone https://github.com/kbaseapps/cb_annotation_ontology_api.git cb_annotation_ontology_api 97f9525aec95390f43f8be88491f35bb11bd589c && \
    pin_clone https://github.com/ModelSEED/ModelSEEDDatabase.git ModelSEEDDatabase f7527bced46ddcc78253883b97bddee4a584daeb && \
    pin_clone https://github.com/ModelSEED/ModelSEEDTemplates.git ModelSEEDTemplates 22942996dacbb518cdd265b0993ca96c93a03ebd

# Install dependency packages in the order they expect.
# cobrakbase first (no deps on others), then ModelSEEDpy, then KBUtilLib.
RUN pip install --no-cache-dir -e /deps/cobrakbase && \
    pip install --no-cache-dir -e /deps/ModelSEEDpy && \
    pip install --no-cache-dir -e /deps/KBUtilLib

WORKDIR /app

# Copy modelseed-api source from build context. With CI, the context is
# this repo's checkout; locally, run from the repo root.
COPY src/ /app/src/
COPY data/ /app/data/
COPY pyproject.toml /app/

# Standalone jobs use subprocesses, so Redis/Celery is intentionally omitted.
# Include MCP because the same image is the documented local AI interface.
RUN pip install --no-cache-dir -e ".[modeling,mcp]"

# KBUtilLib's current dependency-manager fallback searches beside the Python
# installation even when explicit data paths are supplied.
RUN ln -s /deps/cb_annotation_ontology_api /usr/local/lib/python3.11/cb_annotation_ontology_api && \
    ln -s /deps/ModelSEEDDatabase /usr/local/lib/python3.11/ModelSEEDDatabase

# Keep the classifier runtime compatible with ModelSEEDpy's pinned sklearn.
# Then verify reconstruction dependencies resolve and pre-download the classifier.
RUN pip install --no-cache-dir --force-reinstall "numpy<2" "scikit-learn==1.2.0" && \
    python -c "from kbutillib import MSReconstructionUtils; MSReconstructionUtils(config_file=False, token_file=None, kbase_token_file=None, token={'patric': 'unused', 'kbase': 'unused'})" && \
    python -c "from modelseedpy.helpers import get_classifier; get_classifier('knn_ACNP_RAST_filter_01_17_2023')"

# Default configuration.
ENV MODELSEED_MODELSEED_DB_PATH=/deps/ModelSEEDDatabase \
    MODELSEED_TEMPLATES_PATH=/deps/ModelSEEDTemplates/templates/v7.0 \
    MODELSEED_CB_ANNOTATION_ONTOLOGY_API_PATH=/deps/cb_annotation_ontology_api \
    MODELSEED_JOB_STORE_DIR=/tmp/modelseed-jobs \
    MODELSEED_HOST=0.0.0.0 \
    MODELSEED_PORT=8000 \
    MODELSEED_STORAGE_BACKEND=local \
    MODELSEED_LOCAL_DATA_DIR=/data/modelseed
# Local-storage by default: no PATRIC account needed, models persist
# under MODELSEED_LOCAL_DATA_DIR. Users who want PATRIC-workspace mode
# can override MODELSEED_STORAGE_BACKEND=workspace at run time.

# WORKAROUND: cobrakbase.KBaseAPI() reads a token from ~/.kbase/token
# even when not connecting to KBase. Required by MSReconstructionUtils
# init. Dummy value is fine.
ENV KB_AUTH_TOKEN=unused
RUN mkdir -p /root/.kbase && echo "unused" > /root/.kbase/token

# Make sure the default local-data dir exists so first-run writes work
# without the user pre-creating + bind-mounting it.
RUN mkdir -p /data/modelseed && chmod 777 /data/modelseed
VOLUME /data/modelseed

EXPOSE 8000

WORKDIR /app/src
CMD ["python", "-m", "uvicorn", "modelseed_api.main:app", "--host", "0.0.0.0", "--port", "8000"]