docker-compose.yml

version: '3.8'

services:
  # ─────────────────────────────────────────────
  # FastAPI Backend
  # ─────────────────────────────────────────────
  api:
    build: .
    image: ai_gallery
    container_name: ai_engine
    ports:
      - "8000:8000"
    volumes:
      - .:/app                              # Hot-reload for development
      - model_cache:/root/.cache/huggingface # Persist downloaded models
    environment:
      - QDRANT_HOST=qdrant_db
      - QDRANT_PORT=6333
      - QDRANT_COLLECTION=video_embeddings
      # Tell qdrant.py to use HTTP server mode (not embedded local)
      - QDRANT_USE_SERVER=true
    depends_on:
      qdrant_db:
        condition: service_healthy

  # ─────────────────────────────────────────────
  # Qdrant Vector Database
  #
  # Speed-first tuning (speed > RAM):
  #   • async_scorer=true     → parallel scoring threads (FAST, default ON)
  #   • mmap_threshold_kb     → only raw FP32 vecs are on disk; INT8 index
  #                             and HNSW graph are kept in RAM by the client
  #                             config (always_ram=True / on_disk=False).
  #   • max_search_threads=0  → let Qdrant use all CPU cores for search
  # ─────────────────────────────────────────────
  qdrant_db:
    image: qdrant/qdrant:latest
    ports:
      - "6333:6333"   # REST API
      - "6334:6334"   # gRPC (faster bulk ops; enable prefer_grpc in client)
    volumes:
      - qdrant_data:/qdrant/storage
    environment:
      # ── Segment storage ────────────────────────────────────────────────
      # Segments whose on-disk raw-vector size exceeds this threshold are
      # mmap-ed from disk.  50 MB is generous — only raw FP32 vecs hit disk;
      # the INT8 index is kept in RAM via always_ram=True in the collection.
      - QDRANT__STORAGE__MMAP_THRESHOLD_KB=51200

      # ── Async parallel scorer (CRITICAL for speed) ────────────────────
      # false was the old value — it serialised ALL scoring. Flip to true.
      # This allows Qdrant to score candidates on multiple threads.
      - QDRANT__STORAGE__ASYNC_SCORER=true

      # ── Search thread pool ─────────────────────────────────────────────
      # 0 = use all available CPU cores.  Caps at the number of vCPUs.
      - QDRANT__STORAGE__MAX_SEARCH_THREADS=0

      # ── Optimizer (indexing) ───────────────────────────────────────────
      # How many segments to merge concurrently during background optimisation.
      - QDRANT__STORAGE__OPTIMIZERS__MAX_OPTIMIZATION_THREADS=2

      # ── WAL (Write-Ahead Log) ──────────────────────────────────────────
      # Larger WAL capacity = fewer flushes = faster ingestion.
      - QDRANT__STORAGE__WAL__WAL_CAPACITY_MB=64

      # ── Payload storage ────────────────────────────────────────────────
      # Keep payload on disk (it is tiny and rarely read at search time).
      - QDRANT__STORAGE__ON_DISK_PAYLOAD=true

      # ── Telemetry ──────────────────────────────────────────────────────
      - QDRANT__TELEMETRY_DISABLED=true
    healthcheck:
      test: ["CMD-SHELL", "wget -qO- http://localhost:6333/healthz || exit 1"]
      interval: 10s
      timeout: 10s
      retries: 10
      start_period: 30s

volumes:
  qdrant_data:
  model_cache: