one api

Chinese | English | Japanese

Deploy to Zeabur

Zeabur's server is abroad, which automatically solves the network problem, and the free amount is enough for personal use.

1. First, fork a copy of the code.
2. Go to Zeabur, log in, and enter the console.
3. Create a new Project, select Marketplace in Service -> Add Service, select MySQL, and record the connection parameters (user name, password, address, port).
4. Copy the link parameters and run create database create database `one-api` .
5. Then in Service -> Add Service, select Git (authorization is required for the first use), and select your fork repository.
6. Deploy will start automatically and cancel first. Go to Variable below, add a PORT with a value of 3000 , add a SQL_DSN with a value of <username>:<password>@tcp(<addr>:<port>)/one-api , and save. Note that if SQL_DSN is not filled in, the data will not be persisted and the data will be lost after redeployment.
7. Select Redeploy.
8. Go to Domains below and select a suitable domain name prefix, such as "my-one-api", and the final domain name is "my-one-api.zeabur.app", or you can CNAME your own domain name.
9. Wait for the deployment to complete, click the generated domain name to enter the One API.

Deploy to Render

Render provides free credit limits, which can further increase the credit limit after binding the card.

Render can directly deploy docker images without the need fork repository: https://dashboard.render.com

Configuration

The system itself is out of the box.

You can configure it by setting environment variables or command line parameters.

After the system starts, use root user to log in to the system and make further configuration.

Note : If you don't know the meaning of a configuration item, you can temporarily delete the value to see further prompt text.

How to use

Add your API Key to渠道page, and then add a new access token to令牌page.

Then you can use your token to access the One API, which is the same as the OpenAI API.

You need to set the API Base as the deployment address of your One API in various places where OpenAI API is used, for example: https://openai.justsong.cn , and API Key is the token you generate in the One API.

Note that the specific API Base format depends on the client you are using.

For example, the official library of OpenAI:

OPENAI_API_KEY= " sk-xxxxxx "
OPENAI_API_BASE= " https://<HOST>:<PORT>/v1 " 

 graph LR
    A(User)
    A --->|Request with key distributed by One API|B(One API)
    B -->|Relay request|C(OpenAI)
    B -->|Relay request|D(Azure)
    B -->|Relay request|E (other downstream channels of OpenAI API format)
    B -->|Relay and modify the request body and return body|F (downstream channel in non-OpenAI API format)

You can specify which channel to use to handle this request by adding the channel ID after the token, for example: Authorization: Bearer ONE_API_KEY-CHANNEL_ID . Note that the token created by the administrator user must be required to specify the channel ID.

If not added, multiple channels will be used in load balancing.

Environment variables

The One API supports reading environment variables from .env files. Please refer to .env.example file. Please rename it to .env when using it.

1. REDIS_CONN_STRING : Redis will be used as cache after setting.
   • Example: REDIS_CONN_STRING=redis://default:redispw@localhost:49153
   • If the database access latency is very low, there is no need to enable Redis, and data lag will occur after enabling it.
2. SESSION_SECRET : After setting, a fixed session key will be used, so that the cookies of the logged-in user will still be valid after the system restarts.
   • Example: SESSION_SECRET=random_string
3. SQL_DSN : The specified database will be used instead of SQLite after setting, use MySQL or PostgreSQL.
   • example:
     • MySQL: SQL_DSN=root:123456@tcp(localhost:3306)/oneapi
     • PostgreSQL: SQL_DSN=postgres://postgres:123456@localhost:5432/oneapi (Adaptation, feedback is welcome)
   • Note that you need to establish the database oneapi in advance. There is no need to manually create tables, and the program will automatically create tables.
   • If using a local database: the deployment command can be added --network="host" so that programs in the container can access MySQL on the host.
   • If using cloud database: If the cloud server needs to verify the identity, you need to add ?tls=skip-verify in the connection parameters.
   • Please modify the following parameters according to your database configuration (or keep the default value):
     • SQL_MAX_IDLE_CONNS : The maximum number of idle connections, default is 100 .
     • SQL_MAX_OPEN_CONNS : The maximum number of open connections, default is 1000 .
       • If an error Error 1040: Too many connections , please reduce the value appropriately.
     • SQL_CONN_MAX_LIFETIME : The maximum life cycle of a connection, default is 60 , in minutes.
4. LOG_SQL_DSN : After setting, a separate database will be used for the logs table. Please use MySQL or PostgreSQL.
5. FRONTEND_BASE_URL : After setting, the redirect page request will be requested to the specified address, only from the server settings.
   • Example: FRONTEND_BASE_URL=https://openai.justsong.cn
6. MEMORY_CACHE_ENABLED : Enable memory caching will cause a certain delay in updating the user quota. The optional values ​​are true and false , and if not set, the default is false .
   • Example: MEMORY_CACHE_ENABLED=true
7. SYNC_FREQUENCY : The frequency of configurations synchronized with the database with cache enabled, in seconds, default is 600 seconds.
   • Example: SYNC_FREQUENCY=60
8. NODE_TYPE : After setting, the node type will be specified. The optional values ​​are master and slave . If not set, the default is master .
   • Example: NODE_TYPE=slave
9. CHANNEL_UPDATE_FREQUENCY : After setting, the channel balance will be updated regularly, in minutes, and if it is not set, it will not be updated.
   • Example: CHANNEL_UPDATE_FREQUENCY=1440
10. CHANNEL_TEST_FREQUENCY : After setting, the channel will be checked regularly, the unit is minutes, and if it is not set, the inspection will not be carried out. +Example: CHANNEL_TEST_FREQUENCY=1440
11. POLLING_INTERVAL : The request interval when batch update channel balance and testing availability, in seconds, with no interval by default.
    • Example: POLLING_INTERVAL=5
12. BATCH_UPDATE_ENABLED : Enable database batch update aggregation, which will cause a certain delay in the update of user quota. The optional values ​​are true and false , and if not set, the default is false .
    • Example: BATCH_UPDATE_ENABLED=true
    • If you are having too many database connections, you can try enabling this option.
13. BATCH_UPDATE_INTERVAL=5 : The time interval for batch updates the aggregate, in seconds, default is 5 .
    • Example: BATCH_UPDATE_INTERVAL=5
14. Request frequency limit:
    • GLOBAL_API_RATE_LIMIT : Global API rate limit (except relay requests), the maximum number of requests within three minutes of a single ip, default is 180 .
    • GLOBAL_WEB_RATE_LIMIT : Global Web rate limit, the maximum number of requests within three minutes of a single IP, default is 60 .
15. Encoder cache settings:
    • TIKTOKEN_CACHE_DIR : When the program starts, some common word encodings will be downloaded online, such as: gpt-3.5-turbo . In some network environments, it may cause startup problems. You can configure this directory to cache data and migrate to offline environments.
    • DATA_GYM_CACHE_DIR : Currently, this configuration function is consistent with TIKTOKEN_CACHE_DIR , but the priority is not as high as it is.
16. RELAY_TIMEOUT : The relay timeout setting is set in seconds, and the timeout time is not set by default.
17. RELAY_PROXY : Use this proxy to request the API after setting.
18. USER_CONTENT_REQUEST_TIMEOUT : The timeout for downloading content uploaded by the user, in seconds.
19. USER_CONTENT_REQUEST_PROXY : After setting, use this proxy to request content uploaded by the user, such as images.
20. SQLITE_BUSY_TIMEOUT : SQLite lock waits for timeout setting, unit is milliseconds, default is 3000 .
21. GEMINI_SAFETY_SETTING : The security setting of Gemini, default BLOCK_NONE .
22. GEMINI_VERSION : The Gemini version used by One API, defaults to v1 .
23. THEME : The theme settings of the system, default is default , please refer to the specific optional values ​​here.
24. ENABLE_METRIC : Whether to disable the channel according to the request success rate, it is not enabled by default, and the optional values ​​are true and false .
25. METRIC_QUEUE_SIZE : The request success rate statistics queue size, default is 10 .
26. METRIC_SUCCESS_RATE_THRESHOLD : The request success rate threshold is 0.8 by default.
27. INITIAL_ROOT_TOKEN : If this value is set, a root user token with the value of the environment variable will be automatically created when the system is first booted.
28. INITIAL_ROOT_ACCESS_TOKEN : If this value is set, a system management token is automatically created when the system is first started.
29. ENFORCE_INCLUDE_USAGE : Whether to force use to return in the stream model, it is not enabled by default, optional values ​​are true and false .

Command line parameters

1. --port <port_number> : Specifies the port number that the server listens to, the default is 3000 .
   • Example: --port 3000
2. --log-dir <log_dir> : Specify the log folder. If it is not set, it is saved to the logs folder in the working directory by default.
   • Example: --log-dir ./logs
3. --version : Print the system version number and exit.
4. --help : Check the command usage help and parameter description.

Demo

Online Demo

Note that this demonstration site does not provide external services: https://openai.justsong.cn

Screenshot display

Frequently Asked Questions

1. What is the credit limit? How to calculate it? Is there any problem with the quota calculation of One API?
   • Credit = Grouping magnification* Model magnification* (Tip token number + complete token number* Completion magnification)
   • The completion ratio is fixed to 1.33 for GPT3.5 and 2 for GPT4, which is consistent with the official.
   • If it is non-stream mode, the official interface will return the total token consumed, but you should pay attention to the different consumption rates of prompts and completion.
   • Note that the default magnification of One API is the official magnification, which has been adjusted.
2. Why is the account quota insufficient?
   • Please check whether your token limit is sufficient, this is separate from the account limit.
   • The token limit is only for the user to set the maximum usage, and users can set it freely.
3. Tip: No channels available?
   • Please check the user grouping and channel grouping settings.
   • and the model settings of the channel.
4. Channel test error: invalid character '<' looking for beginning of value
   • This is because the return value is not a legal JSON, but an HTML page.
   • It is highly likely that the IP or proxy node of your deployment site has been banned by CloudFlare.
5. ChatGPT Next Web Error: Failed to fetch
   • Do not set BASE_URL when deploying.
   • Check whether your interface address and API Key are filled in correctly.
   • Check whether HTTPS is enabled, and the browser will intercept HTTP requests under the HTTPS domain name.
6. Error:当前分组负载已饱和，请稍后再试
   • Upstream channel 429.
7. Will my data be lost after upgrading?
   • If using MySQL, it won't.
   • If using SQLite, you need to mount the volume persisted one-api.db database file according to the deployment command I gave, otherwise the data will be lost after the container is restarted.
8. Do I need to make changes to the database before upgrading?
   • Generally, it is not necessary, and the system will automatically adjust it during initialization.
   • If necessary, I will explain it in the update log and give the script.
9. An error was reported after manually modifying the database:数据库一致性已被破坏，请联系管理员?
   • This is the channel id that detects that some records in the ability table do not exist. This is most likely because you deleted the records in the channel table but did not synchronize the invalid channels in the ability table to clean up the invalid channels.
   • For each channel, the supported models need to have a record of a special ability table that indicates that the channel supports the model.

Related projects

• FastGPT: Knowledge Base Q&A System Based on LLM Large Language Model
• ChatGPT Next Web: Have your own cross-platform ChatGPT application with one click
• VChart: Not only is it a multi-end chart library out of the box, it is also a vivid and flexible data storyteller.
• VMind: Not only is it automatic, it is also very smart. Open source intelligent visualization solution.

Notice

This project is open source using the MIT protocol. On this basis , the attribution and links to this project must be retained at the bottom of the page. If you do not want to retain the attribution, you must first obtain authorization.

It is also applicable to second-department projects based on this project.

According to the MIT agreement, users must bear the risks and responsibilities of using this project at their own expense, and the developers of this open source project have nothing to do with this.