aiproxy

README

English | 简体中文

AI Proxy

Next-generation AI gateway, using OpenAI as the protocol entry point.

Feature

• Intelligent error retry
• Channel selection based on priority and error rate
• Alert notifications
  • Channel balance warning
  • Error rate warning
  • Unauthorized channel warning
  • and more...
• Logging and auditing
  • Comprehensive request log data
  • Request and response body recording
  • Request log tracing
• Data statistics and analysis
  • Request volume statistics
  • Error volume statistics
  • RPM TPM statistics
  • Consumption statistics
  • Model statistics
  • Channel error rate analysis
  • and more...
• Rerank support
• PDF support
• STT model mapping support
• Multi-tenant system separation
• Model RPM TPM limits
• Think model support <think> split to reasoning_content
• Prompt Token Cache billing support
• Inline tiktoken, no need to download tiktoken file
• API Swagger documentation support http://host:port/swagger/index.html

How to use

Sealos

Use Sealos built-in model capabilities, click to Sealos.

FastGPT

Use AI Proxy to access models, click to FastGPT.

Deploy

Use Docker

docker run -d --name aiproxy -p 3000:3000 -v $(pwd)/aiproxy:/aiproxy ghcr.io/labring/aiproxy:latest

Use Docker Compose

Copy docker-compose.yaml to directory.

docker-compose up -d

Envs

Basic Configuration

• LISTEN: The listen address, default is :3000
• ADMIN_KEY: The admin key for the AI Proxy Service, admin key is used to admin api and relay api, default is empty
• INTERNAL_TOKEN: Internal token for service authentication, default is empty
• FFPROBE_ENABLED: Whether to enable ffprobe, default is false

Debug Options

• DEBUG: Enable debug mode, default is false
• DEBUG_SQL: Enable SQL debugging, default is false

Database Options

• SQL_DSN: The database connection string, default is empty, eg: postgres://postgres:postgres@localhost:5432/postgres
• LOG_SQL_DSN: The log database connection string, default is empty, eg: postgres://postgres:postgres@localhost:5432/postgres
• REDIS_CONN_STRING: The redis connection string, default is empty, eg: redis://localhost:6379
• DISABLE_AUTO_MIGRATE_DB: Disable automatic database migration, default is false
• SQL_MAX_IDLE_CONNS: The maximum number of idle connections in the database, default is 100
• SQL_MAX_OPEN_CONNS: The maximum number of open connections to the database, default is 1000
• SQL_MAX_LIFETIME: The maximum lifetime of a connection in seconds, default is 60
• SQLITE_PATH: The path to the sqlite database, default is aiproxy.db
• SQL_BUSY_TIMEOUT: The busy timeout for the database, default is 3000

Notify Options

• NOTIFY_NOTE: Custom notification note, default is AI Proxy
• NOTIFY_FEISHU_WEBHOOK: The feishu notify webhook url, default is empty, eg: https://open.feishu.cn/open-apis/bot/v2/hook/xxxx

Model Configuration

• DISABLE_MODEL_CONFIG: Disable model configuration, default is false
• RETRY_TIMES: Number of retry attempts, default is 0
• ENABLE_MODEL_ERROR_AUTO_BAN: Enable automatic banning of models with errors, default is false
• MODEL_ERROR_AUTO_BAN_RATE: Rate threshold for auto-banning models with errors, default is 0.3
• TIMEOUT_WITH_MODEL_TYPE: Timeout settings for different model types, default is {}
• DEFAULT_CHANNEL_MODELS: Default models for each channel, default is {}
• DEFAULT_CHANNEL_MODEL_MAPPING: Model mapping for each channel, default is {}

Logging Configuration

• LOG_STORAGE_HOURS: Hours to store logs (0 means unlimited), default is 0
• LOG_CONTENT_STORAGE_HOURS: Hours to store log content ip endpoint ttfb_milliseconds, default is 0
• SAVE_ALL_LOG_DETAIL: Save all log details, default is false
• LOG_DETAIL_REQUEST_BODY_MAX_SIZE: Maximum size for request body in log details, default is 128KB
• LOG_DETAIL_RESPONSE_BODY_MAX_SIZE: Maximum size for response body in log details, default is 128KB
• LOG_DETAIL_STORAGE_HOURS: Hours to store log details, default is 72 (3 days)
• CLEAN_LOG_BATCH_SIZE: Batch size for cleaning logs, cleaning interval is 1 minute, default is 2000

Service Control

• DISABLE_SERVE: Disable serving requests, default false
• GROUP_MAX_TOKEN_NUM: Maximum number of tokens per group (0 means unlimited), default is 0
• GROUP_CONSUME_LEVEL_RATIO: Consumption level ratio for groups, default is {}
• GEMINI_SAFETY_SETTING: Safety setting for Gemini models, default is BLOCK_NONE
• BILLING_ENABLED: Enable billing functionality, default is true
• IP_GROUPS_THRESHOLD: IP group threshold, when the same IP is used by multiple groups, send a warning, default is 0
• IP_GROUPS_BAN_THRESHOLD: IP group ban threshold, when the same IP is used by multiple groups, ban it and all groups, default is 0