Full docker compose for Nvidia DGX Spark GB10 behind a reverse proxy

[gardner]

The reverse proxy handles authentication and the browser extension seems to pick up and use the auth tokens.

services:
  hister:
    image: ghcr.io/asciimoo/hister:latest
    container_name: hister
    user: '1000:1000'
    restart: unless-stopped
    environment:
      HISTER__SERVER__ADDRESS: 0.0.0.0:4433
      HISTER__SERVER__BASE_URL: https://hister.example.com

      HISTER__SEMANTIC_SEARCH__ENABLE: "true"
      HISTER__SEMANTIC_SEARCH__EMBEDDING_ENDPOINT: http://embed:80/v1/embeddings
      HISTER__SEMANTIC_SEARCH__EMBEDDING_MODEL: Qwen/Qwen3-Embedding-0.6B
      HISTER__SEMANTIC_SEARCH__DIMENSIONS: "1024"
      HISTER__SEMANTIC_SEARCH__MAX_CONTEXT_LENGTH: "4096"
      HISTER__SEMANTIC_SEARCH__CHUNK_OVERLAP: "128"
      HISTER__SEMANTIC_SEARCH__QUERY_PREFIX: "Instruct: Given a search query, retrieve relevant passages from the indexed documents that answer the query. Query: "
      HISTER__SEMANTIC_SEARCH__DOCUMENT_PREFIX: ""
      HISTER__SEMANTIC_SEARCH__SIMILARITY_THRESHOLD: "0.1"
      HISTER__SEMANTIC_SEARCH__RESULT_LIMIT: "50"
      HISTER__SEMANTIC_SEARCH__SEMANTIC_WEIGHT: "0.4"
      HISTER__SEMANTIC_SEARCH__MAX_EMBEDDING_CONCURRENCY: "4"
    volumes:
      - ./data:/hister/data
    ports:
      - 4433:4433

  embed:
    image: ghcr.io/huggingface/text-embeddings-inference:121-latest
    environment:
      MODEL_ID: Qwen/Qwen3-Embedding-0.6B
      POOLING: last-token
      MAX_BATCH_TOKENS: "32768"
    volumes:
      - tei-data:/data
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
volumes:
  tei-data:

Please note the 121-latest tag at the end of the TEI image. That's the one that works on Spark GB10. You would need to tweak / adjust it for other Nvidia platforms. You can see a full list of image tags here. You can find out which Compute Capability your GPU supports here. Then just remove the period. So, if you have an RTX 3060 with Compute Capability of 8.6 then the image that would be 86-latest.