How to Use LangChain or Agent Mode

[diskun00]

Hi Marcin,

Thank you for creating this nice app! After worked with Langchain Agent last year, I was always looking forward to set up a local agent in my own laptop. Different from what I built in jupyter notebook, to have a UI for convenient interaction with the agent is really essential. I am so happy to see your project.

While setup agent mode, I find it not clear how to make it work. Could you give an example config? For example, i want to set up an agent

1. which can access browser
2. use python
3. run bash commands

I saw your implemented plugins covered all. But how could i configure it in PyGPT? I used to code an LangChain anget and expose those plugins as functions with description so that agent knows which and how to utilize them.

Besides, I am unable to find setting panel but only see it in the first launch of the app. FYI, I run it using clone repo from M2 MBP. I think maybe if i can access the settings page, i might figure it out the problem as well.

Thank you again!

[szczyglis-dev]

Hi ;)

The agent mode works very similarly here to what you described, but all the configuration for this is in the settings (Plugins -> Settings and Config -> Settings) which you don't see in the menu.

This will be a problem related to MacOS trying to manage menu items with specific names in its own way, when mapping them to a global menu:

https://doc.qt.io/qt-5/qmenubar.html#qmenubar-as-a-global-menu-bar

To resolve this, in the current release (2.0.152), I removed all roles for menu items that are specially treated by MacOS, such as Settings, About, or Exit. This trick should help - please check if both Settings options now appear correctly in the menu and give me a feedback - if this doesn't help, I will try to solve it in another way.

---

How to use Agents:

• Choose the mode for the Agent (Chat is default) in Config -> Settings -> Agent - Agent mode is just a virtual mode which runs another mode (like Chat) in a loop
• Enable the plugins you want to use, configure them (API keys, etc.) in Plugins -> Settings and enable the Execute commands checkbox to enable executing commands from plugins
• Switch into Agent mode, choose the number of iterations (0 = infinity),
• Choose the default system prompt to handle the Agent or create a new one
• Send the input prompt with the goal to achieve - the Agent will work in a loop until the specified goal is reached
• You can also use the plugin Autonomous in any non-Agent mode to run Agent-mode inline (it works similarly to Agent mode, but in standard modes like Chat)

Langchain:

• At this moment, there are not too many wrappers for Langchain - by default, you can use OpenAI or HuggingFace wrappers
• The model used for Langchain must be configured in the Config -> Models dialog - all the options required for the Langchain wrapper (like API key, etc.) must be provided on the list
• All provided options will be passed directly to the Langchain wrapper as their kwargs
• By default, there are configurations only for OpenAI models
• You can also create your own wrapper and register it into the app (see the Managing Models section in Docs)

Hope this helps ;)

[diskun00]

Thank you for quick reply! Now i could see the setting after the pulling the code. Will check other features later.