Skip to content

Docker Basic Deployment

This guide shows how to run Prompt Optimizer with Docker and how to configure custom OpenAI-compatible models with runtime environment variables.

Quick Start

docker run -d -p 8081:80 \
  --restart unless-stopped \
  --name prompt-optimizer \
  linshen/prompt-optimizer:latest

Open the app at http://localhost:8081.

Custom Model Environment Variables

Custom models support the following runtime variables:

VITE_CUSTOM_API_KEY_<suffix>=your-api-key
VITE_CUSTOM_API_BASE_URL_<suffix>=your-base-url
VITE_CUSTOM_API_MODEL_<suffix>=your-model-name
VITE_CUSTOM_API_PARAMS_<suffix>=json-object-string

Notes:

  • KEY, BASE_URL, and MODEL are required
  • PARAMS is optional and must be a JSON object string
  • PARAMS is injected into the final OpenAI-compatible request body
  • reserved keys model, messages, and stream are ignored automatically

Example: Docker Run

docker run -d -p 8081:80 \
  -e VITE_CUSTOM_API_KEY_nvidia=nvapi-xxx \
  -e VITE_CUSTOM_API_BASE_URL_nvidia=https://integrate.api.nvidia.com/v1 \
  -e VITE_CUSTOM_API_MODEL_nvidia=qwen/qwen3.5-397b-a17b \
  -e 'VITE_CUSTOM_API_PARAMS_nvidia={"chat_template_kwargs":{"enable_thinking":true},"temperature":0.6,"top_p":0.95,"max_tokens":16384}' \
  --restart unless-stopped \
  --name prompt-optimizer \
  linshen/prompt-optimizer:latest

This is useful for:

  • standard OpenAI-compatible fields such as temperature, top_p, and max_tokens
  • vendor-specific fields such as NVIDIA NIM's chat_template_kwargs
  • stable Docker defaults that should survive browser cache resets

Example: Docker Compose

services:
  prompt-optimizer:
    image: linshen/prompt-optimizer:latest
    ports:
      - "8081:80"
    restart: unless-stopped
    environment:
      VITE_CUSTOM_API_KEY_nvidia: nvapi-xxx
      VITE_CUSTOM_API_BASE_URL_nvidia: https://integrate.api.nvidia.com/v1
      VITE_CUSTOM_API_MODEL_nvidia: qwen/qwen3.5-397b-a17b
      VITE_CUSTOM_API_PARAMS_nvidia: '{"chat_template_kwargs":{"enable_thinking":true},"temperature":0.6,"top_p":0.95,"max_tokens":16384}'

Install as a mobile app (PWA)

Without ACCESS_PASSWORD, the production app can be installed from trusted HTTPS or browser-trusted localhost. A plain LAN HTTP IP does not meet the installation requirement. Offline use supports reading and editing saved local workspaces; model requests still need a network connection.

Docker permits caching public build resources while keeping runtime config.js, API/model responses and the worker script sw.js under no-store. Updates activate only after choosing Save and update. Preserve these cache rules when adding a reverse proxy.

With ACCESS_PASSWORD, authenticated online use remains available. The protected page does not support offline-shell installation and does not show the install reminder. Do not expose a protected document or cache credential-bearing runtime configuration to enable installation.

Verification

  1. Start the container.
  2. Select the custom model in the UI.
  3. Send a message.
  4. Inspect the request payload in browser DevTools and confirm the extra fields are present.

Troubleshooting

  • If PARAMS is not valid JSON, the model still loads, but the extra parameters are ignored.
  • If the custom model does not appear, make sure all required variables use the same suffix.
  • For complex JSON values in shell commands or Compose files, wrap the entire string in single quotes.