Dockerfile

# syntax=docker/dockerfile:1.7
#
# Eolas backend container — FastAPI + LanceDB + fastembed.
#
# Strategy:
#  - Python 3.11-slim (small base; CPython is enough — onnxruntime brings perf)
#  - Install pip deps in their own layer so source edits don't bust the cache
#  - Pre-warm the BGE ONNX model at build time → first request has no 120 MB download
#  - Ship the LanceDB index inside the image (31 MB) → no volume needed in v0.3.0
#  - Run uvicorn on $PORT (Fly sets this; defaults to 8080)
#
FROM python:3.11-slim AS runtime

WORKDIR /app

# Runtime libs onnxruntime needs (libgomp1 is the OpenMP runtime; the rest is
# wheels). curl is here for the container HEALTHCHECK only.
RUN apt-get update && apt-get install -y --no-install-recommends \
        libgomp1 \
        curl \
    && rm -rf /var/lib/apt/lists/*

# ── Deps layer (cached unless pyproject.toml changes) ────────────────────────
COPY pyproject.toml ./
RUN pip install --no-cache-dir --upgrade pip \
 && pip install --no-cache-dir \
        "httpx>=0.27" \
        "trafilatura>=1.12" \
        "pymupdf>=1.24" \
        "beautifulsoup4>=4.12" \
        "lxml>=5.3" \
        "tiktoken>=0.7" \
        "lancedb>=0.13" \
        "fastembed>=0.4" \
        "rank-bm25>=0.2" \
        "flashrank>=0.2" \
        "fastapi>=0.111" \
        "uvicorn[standard]>=0.30" \
        "python-dotenv>=1.0" \
        "sse-starlette>=2.1" \
        "slowapi>=0.1.9" \
        "google-genai>=0.3" \
        "groq>=0.11" \
        "openai>=1.40" \
        "anthropic>=0.34"

# ── Pre-warm the embedder + reranker models ─────────────────────────────────
# bge-small-en-v1.5: ~120 MB ONNX. ms-marco-MiniLM-L-12-v2: ~22 MB ONNX.
# Baked into the image so cold starts have no network fetch.
RUN python -c "from fastembed import TextEmbedding; TextEmbedding('BAAI/bge-small-en-v1.5')"
RUN python -c "from flashrank import Ranker; Ranker(model_name='ms-marco-MiniLM-L-12-v2')"

# ── Application code + indexed corpus ────────────────────────────────────────
COPY app/ ./app/
COPY widget/ ./widget/
COPY data/lancedb/ ./data/lancedb/

# Runtime configuration
ENV PYTHONUNBUFFERED=1 \
    LOG_FORMAT=json \
    LOG_LEVEL=INFO \
    PORT=8080 \
    LANCEDB_PATH=/app/data/lancedb \
    LANCEDB_TABLE=chunks

EXPOSE 8080

HEALTHCHECK --interval=30s --timeout=10s --start-period=40s --retries=3 \
    CMD curl -fsS http://127.0.0.1:${PORT:-8080}/health || exit 1

# Single worker — the BGE model isn't cheap to load and one process saturates
# a 1-core VM. Scale horizontally on Fly instead of forking workers.
CMD ["sh", "-c", "uvicorn app.main:app --host 0.0.0.0 --port ${PORT:-8080} --workers 1"]