Skip to content

Latest commit

 

History

History
43 lines (27 loc) · 3.03 KB

configure_agent_and_tools.md

File metadata and controls

43 lines (27 loc) · 3.03 KB
Raw

Configure your agent and tools

Steps to complete

Agent configuration

The agent and action plans can be configured in agent.yml.

LLMs

default_llm & default_fast_llm: Set the name of the LLMs you want to use. In llm.py you can add your own model (any model compatible with LangChain e.g. Google, Anthropic, or open source like Llama2).

We currently have 2 ways to choose if default_llm or default_fast_llm is used.

  • If you set force_gpt4 to True in the LLM call in a tool, default_llm will always be used
  • [TO BE CHANGED] Otherwise, default_fast_llm will be used for prompts < 2500 tokens (configurable in ExtendedBaseTool.py currently, needs to be cleaned up to add as a key setting) and default_llm for >=2500 tokens

Tools and Action Plans

Add all the tools in use in tools. Ensure the names match the tool names in tools.py and your custom tools.

Configure the Action Plans available for the Meta Agent to choose from in action_plans. Give each Action Plan a clear description of what the use case is, this will improve the reliability and accuracy of the Meta Agent. Add all the tools in actions. Each sublist is 1 action step, so add tools as subitems if you want to execute them in parallel.

Meta agent prompts

It is very important to have a clear system prompt in system_context for the Meta Agent so that it chooses the right Action Plans (prompt_message typically can typically be kept the same). Always include a role for the agent ("You are an expert in ...") and a clear goal ("Your goal is to select the right action plan.."). Include some principles to ensure the agent has the right behaviour for the use case, e.g. only run an optimization when the agent is very sure the user wants this, as it takes a lot of time. If there are common failure modes in the agent's routing choices, add a principle or add an example of good behaviour to solve it.

Tools configuration

Using a library tool

Check out the library of commonly used tools in services/chat_agent/tools/library. Using these tools is simple; ensure the tool is in the tools list in agent.yml and configure the prompts in tools.yml. Detailed documentation on the library tools can be found in docs/library_tool_docs.

Add a tool

  1. Add your own tool folder to services/chat_agent/tools with a new file yourtool.py
  2. Implement your tool. See template_tool.py for a template tool, or look at the other library tools for inspiration
  3. In tools.py, add the tool in all_tool_classes and import it
  4. Add the tool in tools in agent.yml and add the tool to the applicable action plans
  5. Add your tool and configure the tool and prompts in tools.yml

Optional: Customize the Actions for the tool in the UI, see the UI documentation.