SillyTavern

SillyTavern (or ST for short) is a locally installed user interface that allows you to interact with text generation LLMs, image generation engines, and TTS voice models. Our goal is to empower users with as much utility and control over their LLM prompts as possible, embracing the steep learning curve as part of the fun.

部署

compose.yml 配置如下

services:
  sillytavern:
    build: ..
    container_name: sillytavern
    hostname: sillytavern
    image: ghcr.io/sillytavern/sillytavern:latest
    environment:
      - NODE_ENV=production
      - FORCE_COLOR=1
    ports:
      - '8000:8000'
    volumes:
      - /root/sillytavern/config:/home/node/app/config
      - /root/sillytavern/data:/home/node/app/data
      - /root/sillytavern/plugins:/home/node/app/plugins
      - /root/sillytavern/extensions:/home/node/app/public/scripts/extensions/third-party
    restart: unless-stopped

config.yaml

# -- DATA CONFIGURATION --
# Root directory for user data storage
dataRoot: ./data
# -- SERVER CONFIGURATION --
# Listen for incoming connections
listen: true
# Listen on a specific address, supports IPv4 and IPv6
listenAddress:
  ipv4: 0.0.0.0
  ipv6: '[::]'
# Enables IPv6 and/or IPv4 protocols. Need to have at least one enabled!
# - Use option "auto" to automatically detect support
# - Use true or false (no qoutes) to enable or disable each protocol
protocol:
  ipv4: true
  ipv6: false
# Prefers IPv6 for DNS. Enable this on ISPs that don't have issues with IPv6
dnsPreferIPv6: false
# The hostname that autorun opens.
# - Use "auto" to let the server decide
# - Use options like 'localhost', 'st.example.com'
autorunHostname: 'auto'
# Server port
port: 8000
# Overrides the port for autorun in browser.
# - Use -1 to use the server port.
# - Specify a port to override the default.
autorunPortOverride: -1
# -- SSL options --
ssl:
  enabled: false
  certPath: './certs/cert.pem'
  keyPath: './certs/privkey.pem'
# -- SECURITY CONFIGURATION --
# Toggle whitelist mode
whitelistMode: false
# Whitelist will also verify IP in X-Forwarded-For / X-Real-IP headers
enableForwardedWhitelist: true
# Whitelist of allowed IP addresses
whitelist:
  - ::1
  - 127.0.0.1
# Automatically whitelist Docker host and gateway IPs
whitelistDockerHosts: true
# Toggle basic authentication for endpoints
basicAuthMode: true
# Basic authentication credentials
basicAuthUser:
  username: 'username'
  password: 'password'
# Enables CORS proxy middleware
enableCorsProxy: false
# -- REQUEST PROXY CONFIGURATION --
requestProxy:
  # If a proxy is enabled, all outgoing HTTP/HTTPS requests will be routed through it.
  enabled: false
  # Proxy URL. Possible protocols: http, https, socks, socks5, socks4, pac
  url: 'socks5://username:password@example.com:1080'
  # Proxy bypass list. Requests to these hosts won't be routed through the proxy.
  bypass:
    - localhost
    - 127.0.0.1
# Enable multi-user mode
enableUserAccounts: false
# Enable discreet login mode: hides user list on the login screen
enableDiscreetLogin: false
# Enable's authlia based auto login. Only enable this if you
# have setup and installed Authelia as a middle-ware on your
# reverse proxy
# https://www.authelia.com/
# This will use auto login to an account with the same username
# as that used for authlia. (Ensure the username in authlia
# is an exact match with that in sillytavern)
autheliaAuth: false
# If `basicAuthMode` and this are enabled then
# the username and passwords for basic auth are the same as those
# for the individual accounts
perUserBasicAuth: false

# User session timeout *in seconds* (defaults to 24 hours).
## Set to a positive number to expire session after a certain time of inactivity
## Set to 0 to expire session when the browser is closed
## Set to a negative number to disable session expiration
sessionTimeout: -1
# Disable CSRF protection - NOT RECOMMENDED
disableCsrfProtection: false
# Disable startup security checks - NOT RECOMMENDED
securityOverride: false
# -- LOGGING CONFIGURATION --
logging:
  # Enable access logging to access.log file
  # Records new connections with timestamp, IP address and user agent
  enableAccessLog: true
  # Minimum log level to display in the terminal (DEBUG = 0, INFO = 1, WARN = 2, ERROR = 3)
  minLogLevel: 0
# -- RATE LIMITING CONFIGURATION --
rateLimiting:
  # Use X-Real-IP header instead of socket IP for rate limiting
  # Only enable this if you are using a properly configured reverse proxy (like Nginx/traefik/Caddy)
  preferRealIpHeader: false
# -- ADVANCED CONFIGURATION --
# Open the browser automatically
autorun: true
# Avoids using 'localhost' for autorun in auto mode.
# use if you don't have 'localhost' in your hosts file
avoidLocalhost: false

## BACKUP CONFIGURATION
backups:
  # Common settings for all backup types
  common:
    # Number of backups to keep for each chat and settings file
    numberOfBackups: 50
  chat:
    # Enable automatic chat backups
    enabled: true
    # Maximum number of chat backups to keep per user (starting from the most recent). Set to -1 to keep all backups.
    maxTotalBackups: -1
    # Interval in milliseconds to throttle chat backups per user
    throttleInterval: 10000

# THUMBNAILING CONFIGURATION
thumbnails:
  # Enable thumbnail generation
  enabled: true
  # Image format of avatar thumbnails:
  # * "jpg": best compression with adjustable quality, no transparency
  # * "png": preserves transparency but increases filesize by about 100%
  # Changing this only affects new thumbnails. To recreate the old ones, clear out /thumbnails folder in your user data.
  format: 'jpg'
  # JPG thumbnail quality (0-100)
  quality: 95
  # Maximum thumbnail dimensions per type [width, height]
  dimensions: { 'bg': [160, 90], 'avatar': [96, 144] }

# PERFORMANCE-RELATED CONFIGURATION
performance:
  # Enables lazy loading of character cards. Improves performances with large card libraries.
  # May have compatibility issues with some extensions.
  lazyLoadCharacters: false
  # The maximum amount of memory that parsed character cards can use. Set to 0 to disable memory caching.
  memoryCacheCapacity: '100mb'

# Allow secret keys exposure via API
allowKeysExposure: false
# Skip new default content checks
skipContentCheck: false
# Allowed hosts for card downloads
whitelistImportDomains:
  - localhost
  - cdn.discordapp.com
  - files.catbox.moe
  - raw.githubusercontent.com
# API request overrides (for KoboldAI and Text Completion APIs)
## Note: host includes the port number if it's not the default (80 or 443)
## Format is an array of objects:
## - hosts:
##   - example.com
##   headers:
##     Content-Type: application/json
##   - 127.0.0.1:5001
##   headers:
##     User-Agent: "Googlebot/2.1 (+http://www.google.com/bot.html)"
requestOverrides: []

# EXTENSIONS CONFIGURATION
extensions:
  # Enable UI extensions
  enabled: true
  # Automatically update extensions when a release version changes
  autoUpdate: true
  models:
    # Enables automatic model download from HuggingFace
    autoDownload: true
    # Additional models for extensions. Expects model IDs from HuggingFace model hub in ONNX format
    classification: Cohee/distilbert-base-uncased-go-emotions-onnx
    captioning: Xenova/vit-gpt2-image-captioning
    embedding: Cohee/jina-embeddings-v2-base-en
    speechToText: Xenova/whisper-small
    textToSpeech: Xenova/speecht5_tts

# Additional model tokenizers can be downloaded on demand.
# Disabling will fallback to another locally available tokenizer.
enableDownloadableTokenizers: true
# -- OPENAI CONFIGURATION --
# A placeholder message to use in strict prompt post-processing mode when the prompt doesn't start with a user message
promptPlaceholder: '[Start a new chat]'
openai:
  # Will send a random user ID to OpenAI completion API
  randomizeUserId: false
  # If not empty, will add this as a system message to the start of every caption completion prompt
  # Example: "Perform the instructions to the best of your ability.\n" (for LLaVA)
  # Not used in image inlining mode
  captionSystemPrompt: ''
# -- DEEPL TRANSLATION CONFIGURATION --
deepl:
  # Available options: default, more, less, prefer_more, prefer_less
  formality: default
# -- MISTRAL API CONFIGURATION --
mistral:
  # Enables prefilling of the reply with the last assistant message in the prompt
  # CAUTION: The prefix is echoed into the completion. You may want to use regex to trim it out.
  enablePrefix: false
# -- OLLAMA API CONFIGURATION --
ollama:
  # Controls how long the model will stay loaded into memory following the request
  # * -1: Keep the model loaded indefinitely
  # * 0: Unload the model immediately after the request
  # * N (any positive number): Keep the model loaded for N seconds after the request.
  keepAlive: -1
  # Controls the "num_batch" (batch size) parameter of the generation request
  # * -1: Use the default value of the model
  # * N (positive number): Use the specified value. Must be a power of 2, e.g. 128, 256, 512, etc.
  batchSize: -1
# -- ANTHROPIC CLAUDE API CONFIGURATION --
claude:
  # Enables caching of the system prompt (if supported).
  # https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching
  # -- IMPORTANT! --
  # Use only when the prompt before the chat history is static and doesn't change between requests
  # (e.g {{random}} macro or lorebooks not as in-chat injections).
  # Otherwise, you'll just waste money on cache misses.
  enableSystemPromptCache: false
  # Enables caching of the message history at depth (if supported).
  # https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching
  # -- IMPORTANT! --
  # Use with caution. Behavior may be unpredictable and no guarantees can or will be made.
  # Set to an integer to specify the desired depth. 0 (which does NOT include the prefill)
  # should be ideal for most use cases.
  # Any value other than a non-negative integer will be ignored and caching at depth will not be enabled.
  cachingAtDepth: -1
# -- SERVER PLUGIN CONFIGURATION --
enableServerPlugins: false
# Attempt to automatically update server plugins on startup
enableServerPluginsAutoUpdate: true

参考链接

• alist