Optimize Bee agent performance on tool calling by implementing tool abstractions

[mmurad2]

Is your feature request related to a problem? Please describe.
Certain tool names do not have "intuitive" names which may lead an LLM agent to not perform the right tool call.

Describe the solution you'd like
An abstraction on defined core functionalities that serves as an interface between the model and discrete tools. E.g. internet_search or get_weather_forecast

Then, for example, a developer can contribute a google tool or duckduckgo tool or brave tool that slots into internet_search

Describe alternatives you've considered
NA

Additional context
Related issue: Optimizing OpenMeteo #110

[mmurad2]

• It should be easy to rename the tool
• We should have GenAI intuitive names for our OTTB tools
• It should be easy for a tool contributor to use an existing name (like internet search) and guidance on how to create custom tools names

[Tomas2D]

1. Renaming tool is easy

const wikipedia = new Wikipedia()
wikipedia.name = 'call_wikipedia'

[Tomas2D]

A tool in the framework is a class with a name (e.g., WikipediaTool) and a property name (e.g., Wikipedia).
The second one goes to the system prompt. We can just rename the internal property for all existing tools to satisfy requirements.

Wikipedia -> ???
OpenMeteo -> get_weather_forecast
DuckDuckGo -> get_web_search_results
GoogleSearch ->get_web_search_results
...

This, therefore, implies the following:

• the agent can use only one tool from a given "group" (which may or might not be desired).
• the user will not know (from the messages) which tool (real class) is being used (fortunately, the emitter events information about the original class).
• we will "group" similar tools

or

We just let the user change the tool name to the desired one.

Any thoughts? @dakshiagrawal

[Tomas2D]

Example of creating a simple tool from a complex tool

const complexWeatherTool = new OpenMeteoTool();

const simpleWeatherTool = new DynamicTool({
  name: "get_current_weather",
  description: "Retrieves current weather in a given destination.",
  inputSchema: z.object({
    location: z.string().min(1).describe("Example: New York"),
  }),
  async handler(input) {
    const { result } = await complexWeatherTool.run({
      location: { name: input.location },
      start_date: new Date().toISOString(),
      temperature_unit: "celsius",
    });
    return new StringToolOutput(`Current weather is ${result.current!.temperature_2m}°C`);
  },
});