docker-compose.yml

services:
  qdrant:
    image: qdrant/qdrant:v1.18.2
    restart: unless-stopped
    ports:
      # Loopback only: Qdrant has no auth configured — binding to 0.0.0.0 would
      # let anyone on the network read/delete every vector, bypassing markdex.
      - "127.0.0.1:6333:6333"
    volumes:
      - qdrant_storage:/qdrant/storage

  embedder:
    build: ./services/embedder
    restart: unless-stopped
    environment:
      # fp16 is faster here — Apple Silicon (arm64) has native fp16; measured ~5x vs fp32.
      USE_FP16: "true"
      # rerank short topic chunks; halves cross-encoder cost vs the 512 default.
      RERANK_MAX_LENGTH: "256"
      # English MiniLM cross-encoder: ~55x faster on CPU than bge-reranker-v2-m3
      # (0.1s vs 6s for a 24-doc pool). For multilingual/highest quality on a GPU,
      # set BAAI/bge-reranker-v2-m3 instead.
      RERANK_MODEL: cross-encoder/ms-marco-MiniLM-L-6-v2
    volumes:
      # persists the BGE-M3 + reranker models (~4.5GB, downloaded on first start)
      - hf_cache:/root/.cache/huggingface

  app:
    build: .
    restart: unless-stopped
    depends_on:
      - qdrant
      - embedder
    environment:
      QDRANT_URL: http://qdrant:6333
      EMBEDDER_URL: http://embedder:8000
      # Optional: raises GitHub's 60/hr unauthenticated limit and enables private
      # repos for repo ingestion. Passed through from the host env if set.
      GITHUB_TOKEN: ${GITHUB_TOKEN:-}
      # Optional: when set, mutating API routes require "Authorization: Bearer <token>".
      # Unset keeps the API open (single-user default).
      MARKDEX_API_TOKEN: ${MARKDEX_API_TOKEN:-}
      # Optional: TurboQuant scalar quantization on the dense vector for collections
      # created from now on (0=off, 2=near-lossless ~16x). ON_DISK is required for the
      # RAM savings (stores float32 originals on disk). See README "TurboQuant".
      MARKDEX_QUANT_BITS: ${MARKDEX_QUANT_BITS:-0}
      MARKDEX_QUANT_ON_DISK: ${MARKDEX_QUANT_ON_DISK:-false}
    ports:
      - "4334:4334"

volumes:
  qdrant_storage:
  hf_cache: