New Features
- `[Request: OAIClient]` Add a new request plugins for those models which API format is alike OpenAI but have some additional rules like can not support multi system messages or must have strict user-assistant message orders. It's very useful for those local servering models started by local model servering library like [Xinference](https://github.com/xorbitsai/inference).
HOW TO USE:
python
import Agently
agent_factory = (
Agently.AgentFactory(is_debug=True)
.set_settings("current_model", "OAIClient")
Mixtral for example
.set_settings("model.OAIClient.url" , "https://api.mistral.ai/v1")
if you want to use Moonshot Kimi:
.set_settings("model.OAIClient.url", "https://api.moonshot.cn/v1")
set model name
Mixtral model list: https://docs.mistral.ai/platform/endpoints/
.set_settings("model.OAIClient.options", { "model": "open-mistral-7b" })
Moonshot mode list: https://platform.moonshot.cn/docs/pricing#文本生成模型-moonshot-v1
set API-KEY if needed
.set_settings("model.OAIClient.auth.api_key", "")
set proxy if needed
.set_proxy("http://127.0.0.1:7890")
you can also change message rules
.set_settings("model.OAIClient.message_rules", {
"no_multi_system_messages": True, True by default, will combine multi system messages into one
"strict_orders": True, True by default, will transform messages' order into "User-Assistant-User-Assitant" strictly
"no_multi_type_messages": True, True by default, will only allow text messages
})
)
agent = agent_factory.create_agent()
(
agent
.set_role("You love EMOJI very much and try to use EMOJI in every sentence.")
.chat_history([
{ "role": "user", "content": "It's a beautiful day, isn't it?" },
{ "role": "assistant", "content": "Right, shine and bright!☀️" }
])
.input("What do you suggest us to do today?")
use .start("completions") if your model is a completion model
.start("chat")
)
Update
- `[Request: ERNIE]` Add support for parameter `system` in new API reference, now system prompt will be pass to parameter `system` instead of being transformed into one of the user chat message. https://github.com/Maplemx/Agently/commit/dc52bdc9dfe829675b403e478c28464297fbdcd1
- `[Request]` Optimized prompt about multi items in list https://github.com/Maplemx/Agently/commit/9f378c771a99796845dbbe835f0cbac4c9e0271f
Bug Fixed
- `[Request Alias]` Fixed some bugs that cause `.general()` and `.abstract()` not working https://github.com/Maplemx/Agently/commit/5f6dd5e7e14bf46e5b25b4898fc4767c1d5e7829
- `[Agent Component: Segment]` Fixed a bug that cause streaming handler not working https://github.com/Maplemx/Agently/commit/8ad370c531366d52f1c217d424c0ffc74a42f400
- `[Request: ERNIE]` Fixed some quotation marks conflict https://github.com/Maplemx/Agently/commit/fcdcdf04476ac932a6199d20ea63eb4b4d64c408
---
新功能
- `[模型请求插件: OAIClient]` 新增新的模型请求插件`OAIClient`，用于支持开发者请求那些看起来很像OpenAI API格式的模型（但通常会有些和OpenAI API不一致的潜在规则）。这个请求插件也可以用于请求通过类似[Xinference](https://github.com/xorbitsai/inference)这样的本地模型服务化库启动的本地模型服务。
如何使用：
python
import Agently
agent_factory = (
Agently.AgentFactory(is_debug=True)
.set_settings("current_model", "OAIClient")
这里用Mixtral举例
.set_settings("model.OAIClient.url" , "https://api.mistral.ai/v1")
如果你希望使用月之暗面的Kimi可以参考下面这个url
.set_settings("model.OAIClient.url", "https://api.moonshot.cn/v1")
设置你要使用的具体模型
Mixtral支持的模型列表: https://docs.mistral.ai/platform/endpoints/
.set_settings("model.OAIClient.options", { "model": "open-mistral-7b" })
月之暗面支持的模型列表: https://platform.moonshot.cn/docs/pricing#文本生成模型-moonshot-v1
设置API-KEY（如果需要的话，本地模型可能不需要）
.set_settings("model.OAIClient.auth.api_key", "")
设置代理（如果需要的话）
.set_proxy("http://127.0.0.1:7890")
你也可以通过设置变更消息处理的规则
.set_settings("model.OAIClient.message_rules", {
"no_multi_system_messages": True, 默认开，如果有多条system消息，将会被合并成一条
"strict_orders": True, 默认开，将会强制将消息列转换成“用户-助理-用户-助理”顺序
"no_multi_type_messages": True, 默认开，将只保留文本类消息，并且直接将文本值放到content里
})
)
agent = agent_factory.create_agent()
(
agent
.set_role("You love EMOJI very much and try to use EMOJI in every sentence.")
.chat_history([
{ "role": "user", "content": "It's a beautiful day, isn't it?" },
{ "role": "assistant", "content": "Right, shine and bright!☀️" }
])
.input("What do you suggest us to do today?")
使用.start("completions")可以支持补全模型！
.start("chat")
)
更新
- `[模型请求插件: ERNIE]` 添加了在新的API规范中，对`system`参数的直接支持。现在传递给文心一言的system prompt将会直接传递给API接口的`system`参数，而不再转换成用户对话消息了；
- `[请求优化]` 优化了list中可以包含多个item的prompt提示方法。
问题修复
- `[请求指令]` 修复了导致`.general()`和`.abstract()`不能正常工作的问题；
- `[Agent能力插件: Segment]` 修复了导致流式输出过程中，处理器（handler）无法工作的问题；
- `[模型请求插件: ERNIE]` 修复了一些引号冲突的问题。