Dockerfile.lambda

# The API image for AWS Lambda. Same app, same uvicorn command as the Dockerfile; the
# Lambda Web Adapter extension turns each Lambda invocation into a plain HTTP request
# to localhost:8000, so the FastAPI code needs no Lambda-specific entry point.
FROM python:3.11-slim

# The adapter runs as a Lambda extension: anything in /opt/extensions starts alongside
# the function and relays requests to the port below.
COPY --from=public.ecr.aws/awsguru/aws-lambda-adapter:0.9.1 /lambda-adapter /opt/extensions/lambda-adapter

ENV PYTHONUNBUFFERED=1 \
    PYTHONDONTWRITEBYTECODE=1 \
    PIP_NO_CACHE_DIR=1 \
    OMP_NUM_THREADS=1 \
    OPENBLAS_NUM_THREADS=1 \
    MKL_NUM_THREADS=1 \
    AWS_LWA_PORT=8000 \
    AWS_LWA_READINESS_CHECK_PATH=/health \
    # The fine-tuned embedder's base model ships in the image (models/hf below) and is
    # read from there alone: Lambda has no Hugging Face login, and the model is gated.
    HF_HOME=/app/models/hf \
    HF_HUB_OFFLINE=1 \
    TRANSFORMERS_OFFLINE=1

WORKDIR /app

# The fine-tuned retriever, as FootyVision runs it locally: the LoRA adapter loaded by
# sentence-transformers on top of google/embeddinggemma-300m. Same library versions as
# the environment that built the index, so queries land in the same vector space.
# CPU-only torch: the CUDA build is ~2GB larger and Lambda has no GPU.
RUN pip install --upgrade pip \
    && pip install torch==2.6.0 --index-url https://download.pytorch.org/whl/cpu \
    && pip install sentence-transformers==6.0.1 peft==0.20.0 transformers==5.16.1

# The models change far less often than the code, so they come before it: a code change
# rebuilds only the layers below and leaves these 1.3GB cached.
COPY models/hf ./models/hf
COPY models/embeddinggemma-footyvision ./models/embeddinggemma-footyvision
COPY models/talent ./models/talent
COPY models/value ./models/value
COPY models/spatial ./models/spatial

COPY pyproject.toml README.md ./
COPY src ./src
# Editable install: the code resolves models/ relative to its own path (/app/models).
# boto3 reads Secrets Manager (footyvision/aws_secrets.py).
RUN pip install -e . boto3

# No `init-db` here, unlike the Dockerfile: on Lambda this would run on every cold
# start. Migrations run once per deploy instead.
CMD ["uvicorn", "footyvision.api.main:app", "--host", "0.0.0.0", "--port", "8000"]